[ { "instance_id": "27de7ea0-5b42-403c-9497-847e1fcb1480", "table_id_paper": "D16-1019table_5", "caption": "Link prediction results on the test-I, test-II, and test-all sets of FB122 and WN18 (filtered setting).", "row_header_level": 2, "row_headers": [ [ "FB122", "TransE" ], [ "FB122", "TransH" ], [ "FB122", "TransR" ], [ "FB122", "KALE-Trip" ], [ "FB122", "KALE-Pre" ], [ "FB122", "KALE-Joint" ], [ "WN18", "TransE" ], [ "WN18", "TransH" ], [ "WN18", "TransR" ], [ "WN18", "KALE-Trip" ], [ "WN18", "KALE-Pre" ], [ "WN18", "KALE-Joint" ] ], "column_header_level": 2, "column_headers": [ [ "Test-I", "MRR" ], [ "Test-I", "MED" ], [ "Test-I", "HITS@3 (%)" ], [ "Test-I", "HITS@5 (%)" ], [ "Test-I", "HITS@10 (%)" ], [ "Test-II", "MRR" ], [ "Test-II", "MED" ], [ "Test-II", "HITS@3 (%)" ], [ "Test-II", "HITS@5 (%)" ], [ "Test-II", "HITS@10 (%)" ], [ "Test-ALL", "MRR" ], [ "Test-ALL", "MED" ], [ "Test-ALL", "HITS@3 (%)" ], [ "Test-ALL", "HITS@5 (%)" ], [ "Test-ALL", "HITS@10 (%)" ] ], "contents": [ [ "0.296", "13.0", "36.0", "41.5", "48.1", "0.630", "2.0", "77.5", "82.8", "88.4", "0.480", "2.0", "58.9", "64.2", "70.2" ], [ "0.280", "15.0", "33.6", "39.1", "46.4", "0.606", "2.0", "70.1", "75.4", "82.0", "0.460", "3.0", "53.7", "59.1", "66.0" ], [ "0.283", "16.0", "33.4", "39.2", "46.0", "0.499", "2.0", "57.0", "63.2", "70.1", "0.401", "5.0", "46.4", "52.4", "59.3" ], [ "0.299", "10.0", "36.6", "42.9", "50.2", "0.650", "2.0", "79.0", "83.4", "88.7", "0.492", "2.0", "59.9", "65.2", "71.4" ], [ "0.291", "11.0", "35.8", "41.9", "49.8", "0.713", "1.0", "82.9", "86.1", "89.9", "0.523", "2.0", "61.7", "66.2", "71.8" ], [ "0.325", "9.0", "38.4", "44.7", "52.2", "0.684", "1.0", "79.7", "84.1", "89.6", "0.523", "2.0", "61.2", "66.4", "72.8" ], [ "0.306", "3.0", "57.4", "72.3", "80.1", "0.511", "2.0", "87.5", "95.6", "98.7", "0.453", "2.0", "79.1", "89.1", "93.6" ], [ "0.318", "3.0", "61.7", "72.4", "78.2", "0.653", "2.0", "87.1", "91.4", "94.6", "0.560", "2.0", "80.0", "86.1", "90.0" ], [ "0.299", "3.0", "56.1", "66.7", "74.5", "0.597", "2.0", "75.0", "81.7", "88.0", "0.514", "2.0", "69.7", "77.5", "84.3" ], [ "0.322", "3.0", "61.0", "73.9", "80.8", "0.555", "2.0", "90.6", "96.3", "98.8", "0.490", "2.0", "82.3", "90.1", "93.8" ], [ "0.322", "3.0", "60.6", "74.5", "81.1", "0.612", "2.0", "96.4", "98.6", "99.6", "0.532", "2.0", "86.4", "91.9", "94.4" ], [ "0.338", "3.0", "65.5", "76.3", "82.1", "0.787", "1.0", "93.3", "95.4", "97.2", "0.662", "2.0", "85.5", "90.1", "93.0" ] ], "metrics_loc": "column", "metrics_type": "['MRR', 'MED', 'HITS@3 (%)', 'HITS@5 (%)', 'HITS@10 (%)', 'MRR', 'MED', 'HITS@3 (%)', 'HITS@5 (%)', 'HITS@10 (%)', 'MRR', 'MED', 'HITS@3 (%)', 'HITS@5 (%)', 'HITS@10 (%)']", "target_entity": "['KALE-Trip', 'KALE-Pre', 'KALE-Joint']", "table_html": "
Test-I || MRR || 0.296 Test-I || MED || 13.0 Test-I || HITS@3 (%) || 36.0 Test-I || HITS@5 (%) || 41.5 Test-I || HITS@10 (%) || 48.1 Test-II || MRR || 0.630 Test-II || MED || 2.0 Test-II || HITS@3 (%) || 77.5 Test-II || HITS@5 (%) || 82.8 Test-II || HITS@10 (%) || 88.4 Test-ALL || MRR || 0.480 Test-ALL || MED || 2.0 Test-ALL || HITS@3 (%) || 58.9 Test-ALL || HITS@5 (%) || 64.2 Test-ALL || HITS@10 (%) || 70.2
FB122 || TransH 0.280 15.0 33.6 39.1 46.4 0.606 2.0 70.1 75.4 82.0 0.460 3.0 53.7 59.1 66.0
FB122 || TransR 0.283 16.0 33.4 39.2 46.0 0.499 2.0 57.0 63.2 70.1 0.401 5.0 46.4 52.4 59.3
FB122 || KALE-Trip 0.299 10.0 36.6 42.9 50.2 0.650 2.0 79.0 83.4 88.7 0.492 2.0 59.9 65.2 71.4
FB122 || KALE-Pre 0.291 11.0 35.8 41.9 49.8 0.713 1.0 82.9 86.1 89.9 0.523 2.0 61.7 66.2 71.8
FB122 || KALE-Joint 0.325 9.0 38.4 44.7 52.2 0.684 1.0 79.7 84.1 89.6 0.523 2.0 61.2 66.4 72.8
WN18 || TransE 0.306 3.0 57.4 72.3 80.1 0.511 2.0 87.5 95.6 98.7 0.453 2.0 79.1 89.1 93.6
WN18 || TransH 0.318 3.0 61.7 72.4 78.2 0.653 2.0 87.1 91.4 94.6 0.560 2.0 80.0 86.1 90.0
WN18 || TransR 0.299 3.0 56.1 66.7 74.5 0.597 2.0 75.0 81.7 88.0 0.514 2.0 69.7 77.5 84.3
WN18 || KALE-Trip 0.322 3.0 61.0 73.9 80.8 0.555 2.0 90.6 96.3 98.8 0.490 2.0 82.3 90.1 93.8
WN18 || KALE-Pre 0.322 3.0 60.6 74.5 81.1 0.612 2.0 96.4 98.6 99.6 0.532 2.0 86.4 91.9 94.4
WN18 || KALE-Joint 0.338 3.0 65.5 76.3 82.1 0.787 1.0 93.3 95.4 97.2 0.662 2.0 85.5 90.1 93.0
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "D16-1019", "page_no": 8, "dir": "emnlp2016", "description": "Results. Table 5 show the results in the filtered setting. On each dataset we report the metrics on three sets: test-I, test-II, and the whole test set (denoted by test-all). Test-I contains test triples that cannot be directly inferred by performing pure logical inference on the training set, and hence might be intrinsically more difficult for the rules. The remaining test triples (i.e., the directly inferable ones) are included in Test-II. These triples have either been used directly as training instances in KALE-Pre, or encoded explicitly in training ground rules in KALE-Joint, making this set trivial for the rules to some extent. From the results, we can see that in both settings: (i) KALE-Pre and KALE-Joint outperform (or at least perform as well as) the other methods which use triples alone on almost all the test sets, demonstrating the superiority of incorporating logical rules. (ii) On the test-I sets which contain triples beyond the scope of pure logical inference, KALE-Joint performs significantly better than KALE-Pre. On these sets KALE-Joint can still beat all the baselines by a significant margin in most cases, while KALE-Pre can hardly outperform KALE-Trip. It demonstrates the capability of the joint embedding scenario to learn more predictive embeddings, through which we can make better predictions even beyond the scope of pure logical inference. (iii) On the test-II sets which contain directly inferable triples, KALE-Pre can easily beat all the baselines (even KALE-Joint). That means, for triples covered by pure logical inference, it is trivial to improve the performance by directly incorporating them as training instances.", "class_sentence": "[2, 1, 1, 2, 2, 2, 1, 1, 1, 2, 1, 2]", "sentences": "['Results.', 'Table 5 show the results in the filtered setting.', 'On each dataset we report the metrics on three sets: test-I, test-II, and the whole test set (denoted by test-all).', 'Test-I contains test triples that cannot be directly inferred by performing pure logical inference on the training set, and hence might be intrinsically more difficult for the rules.', 'The remaining test triples (i.e., the directly inferable ones) are included in Test-II.', 'These triples have either been used directly as training instances in KALE-Pre, or encoded explicitly in training ground rules in KALE-Joint, making this set trivial for the rules to some extent.', 'From the results, we can see that in both settings: (i) KALE-Pre and KALE-Joint outperform (or at least perform as well as) the other methods which use triples alone on almost all the test sets, demonstrating the superiority of incorporating logical rules.', '(ii) On the test-I sets which contain triples beyond the scope of pure logical inference, KALE-Joint performs significantly better than KALE-Pre.', 'On these sets KALE-Joint can still beat all the baselines by a significant margin in most cases, while KALE-Pre can hardly outperform KALE-Trip.', 'It demonstrates the capability of the joint embedding scenario to learn more predictive embeddings, through which we can make better predictions even beyond the scope of pure logical inference.', '(iii) On the test-II sets which contain directly inferable triples, KALE-Pre can easily beat all the baselines (even KALE-Joint).', 'That means, for triples covered by pure logical inference, it is trivial to improve the performance by directly incorporating them as training instances.']", "header_mention": "[None, None, ['Test-I', 'Test-II', 'Test-ALL'], ['Test-I'], ['Test-II'], ['Test-I', 'Test-II', 'Test-ALL', 'KALE-Pre', 'KALE-Joint'], ['KALE-Pre', 'KALE-Joint'], ['KALE-Joint', 'KALE-Pre'], ['KALE-Joint', 'KALE-Pre', 'KALE-Trip'], None, ['Test-II', 'KALE-Pre', 'KALE-Joint'], None]", "valid": 1, "image_id": "D16-1019-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Link prediction results on the test-I, test-II, and test-all sets of FB122 and WN18 (filtered setting).

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Test-I || MRR || 0.296Test-I || MED || 13.0Test-I || HITS@3 (%) || 36.0Test-I || HITS@5 (%) || 41.5Test-I || HITS@10 (%) || 48.1Test-II || MRR || 0.630Test-II || MED || 2.0Test-II || HITS@3 (%) || 77.5Test-II || HITS@5 (%) || 82.8Test-II || HITS@10 (%) || 88.4Test-ALL || MRR || 0.480Test-ALL || MED || 2.0Test-ALL || HITS@3 (%) || 58.9Test-ALL || HITS@5 (%) || 64.2Test-ALL || HITS@10 (%) || 70.2
FB122 || TransH0.28015.033.639.146.40.6062.070.175.482.00.4603.053.759.166.0
FB122 || TransR0.28316.033.439.246.00.4992.057.063.270.10.4015.046.452.459.3
FB122 || KALE-Trip0.29910.036.642.950.20.6502.079.083.488.70.4922.059.965.271.4
FB122 || KALE-Pre0.29111.035.841.949.80.7131.082.986.189.90.5232.061.766.271.8
FB122 || KALE-Joint0.3259.038.444.752.20.6841.079.784.189.60.5232.061.266.472.8
WN18 || TransE0.3063.057.472.380.10.5112.087.595.698.70.4532.079.189.193.6
WN18 || TransH0.3183.061.772.478.20.6532.087.191.494.60.5602.080.086.190.0
WN18 || TransR0.2993.056.166.774.50.5972.075.081.788.00.5142.069.777.584.3
WN18 || KALE-Trip0.3223.061.073.980.80.5552.090.696.398.80.4902.082.390.193.8
WN18 || KALE-Pre0.3223.060.674.581.10.6122.096.498.699.60.5322.086.491.994.4
WN18 || KALE-Joint0.3383.065.576.382.10.7871.093.395.497.20.6622.085.590.193.0
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Link prediction results on the test-I, test-II, and test-all sets of FB122 and WN18 (filtered setting).}\n\\begin{tabular}{lrrrrrrrrrrrrrrr}\n\\toprule\n & Test-I $\\parallel$ MRR $\\parallel$ 0.296 & Test-I $\\parallel$ MED $\\parallel$ 13.0 & Test-I $\\parallel$ HITS@3 (\\%) $\\parallel$ 36.0 & Test-I $\\parallel$ HITS@5 (\\%) $\\parallel$ 41.5 & Test-I $\\parallel$ HITS@10 (\\%) $\\parallel$ 48.1 & Test-II $\\parallel$ MRR $\\parallel$ 0.630 & Test-II $\\parallel$ MED $\\parallel$ 2.0 & Test-II $\\parallel$ HITS@3 (\\%) $\\parallel$ 77.5 & Test-II $\\parallel$ HITS@5 (\\%) $\\parallel$ 82.8 & Test-II $\\parallel$ HITS@10 (\\%) $\\parallel$ 88.4 & Test-ALL $\\parallel$ MRR $\\parallel$ 0.480 & Test-ALL $\\parallel$ MED $\\parallel$ 2.0 & Test-ALL $\\parallel$ HITS@3 (\\%) $\\parallel$ 58.9 & Test-ALL $\\parallel$ HITS@5 (\\%) $\\parallel$ 64.2 & Test-ALL $\\parallel$ HITS@10 (\\%) $\\parallel$ 70.2 \\\\\n\\midrule\nFB122 $\\parallel$ TransH & 0.28 & 15 & 33.6 & 39.1 & 46.4 & 0.606 & 2 & 70.1 & 75.4 & 82 & 0.46 & 3 & 53.7 & 59.1 & 66 \\\\\nFB122 $\\parallel$ TransR & 0.283 & 16 & 33.4 & 39.2 & 46 & 0.499 & 2 & 57 & 63.2 & 70.1 & 0.401 & 5 & 46.4 & 52.4 & 59.3 \\\\\nFB122 $\\parallel$ KALE-Trip & 0.299 & 10 & 36.6 & 42.9 & 50.2 & 0.65 & 2 & 79 & 83.4 & 88.7 & 0.492 & 2 & 59.9 & 65.2 & 71.4 \\\\\nFB122 $\\parallel$ KALE-Pre & 0.291 & 11 & 35.8 & 41.9 & 49.8 & 0.713 & 1 & 82.9 & 86.1 & 89.9 & 0.523 & 2 & 61.7 & 66.2 & 71.8 \\\\\nFB122 $\\parallel$ KALE-Joint & 0.325 & 9 & 38.4 & 44.7 & 52.2 & 0.684 & 1 & 79.7 & 84.1 & 89.6 & 0.523 & 2 & 61.2 & 66.4 & 72.8 \\\\\nWN18 $\\parallel$ TransE & 0.306 & 3 & 57.4 & 72.3 & 80.1 & 0.511 & 2 & 87.5 & 95.6 & 98.7 & 0.453 & 2 & 79.1 & 89.1 & 93.6 \\\\\nWN18 $\\parallel$ TransH & 0.318 & 3 & 61.7 & 72.4 & 78.2 & 0.653 & 2 & 87.1 & 91.4 & 94.6 & 0.56 & 2 & 80 & 86.1 & 90 \\\\\nWN18 $\\parallel$ TransR & 0.299 & 3 & 56.1 & 66.7 & 74.5 & 0.597 & 2 & 75 & 81.7 & 88 & 0.514 & 2 & 69.7 & 77.5 & 84.3 \\\\\nWN18 $\\parallel$ KALE-Trip & 0.322 & 3 & 61 & 73.9 & 80.8 & 0.555 & 2 & 90.6 & 96.3 & 98.8 & 0.49 & 2 & 82.3 & 90.1 & 93.8 \\\\\nWN18 $\\parallel$ KALE-Pre & 0.322 & 3 & 60.6 & 74.5 & 81.1 & 0.612 & 2 & 96.4 & 98.6 & 99.6 & 0.532 & 2 & 86.4 & 91.9 & 94.4 \\\\\nWN18 $\\parallel$ KALE-Joint & 0.338 & 3 & 65.5 & 76.3 & 82.1 & 0.787 & 1 & 93.3 & 95.4 & 97.2 & 0.662 & 2 & 85.5 & 90.1 & 93 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Link prediction results on the test-I, test-II, and test-all sets of FB122 and WN18 (filtered setting).\n
\n \n Test-I || MRR || 0.296\n \n Test-I || MED || 13.0\n \n Test-I || HITS@3 (%) || 36.0\n \n Test-I || HITS@5 (%) || 41.5\n \n Test-I || HITS@10 (%) || 48.1\n \n Test-II || MRR || 0.630\n \n Test-II || MED || 2.0\n \n Test-II || HITS@3 (%) || 77.5\n \n Test-II || HITS@5 (%) || 82.8\n \n Test-II || HITS@10 (%) || 88.4\n \n Test-ALL || MRR || 0.480\n \n Test-ALL || MED || 2.0\n \n Test-ALL || HITS@3 (%) || 58.9\n \n Test-ALL || HITS@5 (%) || 64.2\n \n Test-ALL || HITS@10 (%) || 70.2\n
\n FB122 || TransH\n \n 0.280\n \n 15.0\n \n 33.6\n \n 39.1\n \n 46.4\n \n 0.606\n \n 2.0\n \n 70.1\n \n 75.4\n \n 82.0\n \n 0.460\n \n 3.0\n \n 53.7\n \n 59.1\n \n 66.0\n
\n FB122 || TransR\n \n 0.283\n \n 16.0\n \n 33.4\n \n 39.2\n \n 46.0\n \n 0.499\n \n 2.0\n \n 57.0\n \n 63.2\n \n 70.1\n \n 0.401\n \n 5.0\n \n 46.4\n \n 52.4\n \n 59.3\n
\n FB122 || KALE-Trip\n \n 0.299\n \n 10.0\n \n 36.6\n \n 42.9\n \n 50.2\n \n 0.650\n \n 2.0\n \n 79.0\n \n 83.4\n \n 88.7\n \n 0.492\n \n 2.0\n \n 59.9\n \n 65.2\n \n 71.4\n
\n FB122 || KALE-Pre\n \n 0.291\n \n 11.0\n \n 35.8\n \n 41.9\n \n 49.8\n \n 0.713\n \n 1.0\n \n 82.9\n \n 86.1\n \n 89.9\n \n 0.523\n \n 2.0\n \n 61.7\n \n 66.2\n \n 71.8\n
\n FB122 || KALE-Joint\n \n 0.325\n \n 9.0\n \n 38.4\n \n 44.7\n \n 52.2\n \n 0.684\n \n 1.0\n \n 79.7\n \n 84.1\n \n 89.6\n \n 0.523\n \n 2.0\n \n 61.2\n \n 66.4\n \n 72.8\n
\n WN18 || TransE\n \n 0.306\n \n 3.0\n \n 57.4\n \n 72.3\n \n 80.1\n \n 0.511\n \n 2.0\n \n 87.5\n \n 95.6\n \n 98.7\n \n 0.453\n \n 2.0\n \n 79.1\n \n 89.1\n \n 93.6\n
\n WN18 || TransH\n \n 0.318\n \n 3.0\n \n 61.7\n \n 72.4\n \n 78.2\n \n 0.653\n \n 2.0\n \n 87.1\n \n 91.4\n \n 94.6\n \n 0.560\n \n 2.0\n \n 80.0\n \n 86.1\n \n 90.0\n
\n WN18 || TransR\n \n 0.299\n \n 3.0\n \n 56.1\n \n 66.7\n \n 74.5\n \n 0.597\n \n 2.0\n \n 75.0\n \n 81.7\n \n 88.0\n \n 0.514\n \n 2.0\n \n 69.7\n \n 77.5\n \n 84.3\n
\n WN18 || KALE-Trip\n \n 0.322\n \n 3.0\n \n 61.0\n \n 73.9\n \n 80.8\n \n 0.555\n \n 2.0\n \n 90.6\n \n 96.3\n \n 98.8\n \n 0.490\n \n 2.0\n \n 82.3\n \n 90.1\n \n 93.8\n
\n WN18 || KALE-Pre\n \n 0.322\n \n 3.0\n \n 60.6\n \n 74.5\n \n 81.1\n \n 0.612\n \n 2.0\n \n 96.4\n \n 98.6\n \n 99.6\n \n 0.532\n \n 2.0\n \n 86.4\n \n 91.9\n \n 94.4\n
\n WN18 || KALE-Joint\n \n 0.338\n \n 3.0\n \n 65.5\n \n 76.3\n \n 82.1\n \n 0.787\n \n 1.0\n \n 93.3\n \n 95.4\n \n 97.2\n \n 0.662\n \n 2.0\n \n 85.5\n \n 90.1\n \n 93.0\n
\n" }, { "instance_id": "30bf42e2-ffc8-4c1c-b42c-3eeb1bfdd849", "table_id_paper": "D16-1031table_3", "caption": "Comparison on validation perplexity. BoW, TDNN and NABS are the baseline neural compression models with different encoders in Rush et al. (2015)", "row_header_level": 2, "row_headers": [ [ "Model", "Bag-of-Word (BoW)" ], [ "Model", "Convolutional (TDNN)" ], [ "Model", "Attention-Based (NABS) (Rush et al.,,2015)" ], [ "Model", "Forced-Attention (FSC)" ], [ "Model", "Auto-encoding (ASC+FSC1)" ] ], "column_header_level": 1, "column_headers": [ [ "Labelled Data" ], [ "Perplexity" ] ], "contents": [ [ "3.8M", "43.6" ], [ "3.8M", "35.9" ], [ "3.8M", "27.1" ], [ "3.8M", "18.6" ], [ "3.8M", "16.6" ] ], "metrics_loc": "column", "metrics_type": "['Labelled Data', 'Perplexity']", "target_entity": "['Auto-encoding (ASC+FSC1)']", "table_html": "
Labelled Data Perplexity
Model || Bag-of-Word (BoW) 3.8M 43.6
Model || Convolutional (TDNN) 3.8M 35.9
Model || Attention-Based (NABS) (Rush et al.,,2015) 3.8M 27.1
Model || Forced-Attention (FSC) 3.8M 18.6
Model || Auto-encoding (ASC+FSC1) 3.8M 16.6
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D16-1031", "page_no": 8, "dir": "emnlp2016", "description": "In Figure 3, we present the validation perplexity to compare the abilities of the three models to learn the compression languages. The ASC+FSC1(red) employs the same dataset for unlabelled and labelled training, while the ASC+FSC2(black) employs the full unlabelled dataset. Here, the joint ASC+FSC1 model obtains better perplexities than the single discriminative FSC model, but there is not much difference between ASC+FSC1 and ASC+FSC2 when the size of the labelled dataset grows. From the perspective of language modelling, the generative ASC model indeed helps the discriminative model learn to generate good summary sentences. Table 3 displays the validation perplexities of the benchmark models, where the joint ASC+FSC1 model trained on the full labelled and unlabelled datasets performs the best on modelling compression languages.", "class_sentence": "[0, 0, 0, 0, 1]", "sentences": "['In Figure 3, we present the validation perplexity to compare the abilities of the three models to learn the compression languages.', 'The ASC+FSC1(red) employs the same dataset for unlabelled and labelled training, while the ASC+FSC2(black) employs the full unlabelled dataset.', 'Here, the joint ASC+FSC1 model obtains better perplexities than the single discriminative FSC model, but there is not much difference between ASC+FSC1 and ASC+FSC2 when the size of the labelled dataset grows.', 'From the perspective of language modelling, the generative ASC model indeed helps the discriminative model learn to generate good summary sentences.', 'Table 3 displays the validation perplexities of the benchmark models, where the joint ASC+FSC1 model trained on the full labelled and unlabelled datasets performs the best on modelling compression languages.']", "header_mention": "[None, None, None, None, ['Auto-encoding (ASC+FSC1)']]", "valid": 1, "image_id": "D16-1031-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison on validation perplexity. BoW, TDNN and NABS are the baseline neural compression models with different encoders in Rush et al. (2015)

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Labelled DataPerplexity
Model || Bag-of-Word (BoW)3.8M43.6
Model || Convolutional (TDNN)3.8M35.9
Model || Attention-Based (NABS) (Rush et al.,,2015)3.8M27.1
Model || Forced-Attention (FSC)3.8M18.6
Model || Auto-encoding (ASC+FSC1)3.8M16.6
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison on validation perplexity. BoW, TDNN and NABS are the baseline neural compression models with different encoders in Rush et al. (2015)}\n\\begin{tabular}{llr}\n\\toprule\n & Labelled Data & Perplexity \\\\\n\\midrule\nModel $\\parallel$ Bag-of-Word (BoW) & 3.8M & 43.6 \\\\\nModel $\\parallel$ Convolutional (TDNN) & 3.8M & 35.9 \\\\\nModel $\\parallel$ Attention-Based (NABS) (Rush et al.,,2015) & 3.8M & 27.1 \\\\\nModel $\\parallel$ Forced-Attention (FSC) & 3.8M & 18.6 \\\\\nModel $\\parallel$ Auto-encoding (ASC+FSC1) & 3.8M & 16.6 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Comparison on validation perplexity. BoW, TDNN and NABS are the baseline neural compression models with different encoders in Rush et al. (2015)\n
\n \n Labelled Data\n \n Perplexity\n
\n Model || Bag-of-Word (BoW)\n \n 3.8M\n \n 43.6\n
\n Model || Convolutional (TDNN)\n \n 3.8M\n \n 35.9\n
\n Model || Attention-Based (NABS) (Rush et al.,,2015)\n \n 3.8M\n \n 27.1\n
\n Model || Forced-Attention (FSC)\n \n 3.8M\n \n 18.6\n
\n Model || Auto-encoding (ASC+FSC1)\n \n 3.8M\n \n 16.6\n
\n" }, { "instance_id": "37501d34-038b-4893-8dba-247471943414", "table_id_paper": "D16-1043table_4", "caption": "Performance on maximally covered datasets.", "row_header_level": 3, "row_headers": [ [ "Source", "Wikipedia", "Text" ], [ "Source", "Google", "Visual" ], [ "Source", "Google", "MM" ], [ "Source", "Bing", "Visual" ], [ "Source", "Bing", "MM" ], [ "Source", "Flickr", "Visual" ], [ "Source", "Flickr", "MM" ], [ "Source", "ImageNet", "Visual" ], [ "Source", "ImageNet", "MM" ], [ "Source", "ESPGame", "Visual" ], [ "Source", "ESPGame", "MM" ] ], "column_header_level": 6, "column_headers": [ [ "Arch.", "AlexNet", "Agg.", "Mean", "Type/Eval", "SL" ], [ "Arch.", "AlexNet", "Agg.", "Mean", "Type/Eval", "MEN" ], [ "Arch.", "AlexNet", "Agg.", "Max", "Type/Eval", "SL" ], [ "Arch.", "AlexNet", "Agg.", "Max", "Type/Eval", "MEN" ], [ "Arch.", "GoogLeNet", "Agg.", "Mean", "Type/Eval", "SL" ], [ "Arch.", "GoogLeNet", "Agg.", "Mean", "Type/Eval", "MEN" ], [ "Arch.", "GoogLeNet", "Agg.", "Max", "Type/Eval", "SL" ], [ "Arch.", "GoogLeNet", "Agg.", "Max", "Type/Eval", "MEN" ], [ "Arch.", "VGGNet", "Agg.", "Mean", "Type/Eval", "SL" ], [ "Arch.", "VGGNet", "Agg.", "Mean", "Type/Eval", "MEN" ], [ "Arch.", "VGGNet", "Agg.", "Max", "Type/Eval", "SL" ], [ "Arch.", "VGGNet", "Agg.", "Max", "Type/Eval", "MEN" ] ], "contents": [ [ "0.310", "0.682", "0.310", "0.682", "0.310", "0.682", "0.310", "0.682", "0.310", "0.682", "0.310", "0.682" ], [ "0.340", "0.503", "0.334", "0.513", "0.358", "0.495", "0.367", "0.501", "0.342", "0.512", "0.332", "0.494" ], [ "0.380", "0.711", "0.370", "0.719", "0.379", "0.711", "0.365", "0.716", "0.380", "0.714", "0.365", "0.716" ], [ "0.325", "0.567", "0.316", "0.554", "0.310", "0.526", "0.303", "0.520", "0.304", "0.551", "0.289", "0.507" ], [ "0.373", "0.727", "0.360", "0.725", "0.364", "0.723", "0.350", "0.724", "0.361", "0.727", "0.349", "0.719" ], [ "0.234", "0.483", "0.224", "0.441", "0.238", "0.407", "0.236", "0.385", "0.243", "0.460", "0.226", "0.385" ], [ "0.350", "0.715", "0.343", "0.711", "0.347", "0.689", "0.344", "0.703", "0.354", "0.702", "0.339", "0.696" ], [ "0.313", "0.561", "0.313", "0.561", "0.341", "0.540", "0.411", "0.603", "0.404", "0.584", "0.401", "0.578" ], [ "0.362", "0.713", "0.362", "0.713", "0.373", "0.719", "0.401", "0.731", "0.427", "0.727", "0.412", "0.723" ], [ "0.018", "0.448", "0.026", "0.376", "0.063", "0.487", "0.050", "0.434", "0.125", "0.506", "0.106", "0.451" ], [ "0.208", "0.686", "0.187", "0.672", "0.243", "0.700", "0.246", "0.696", "0.269", "0.708", "0.260", "0.698" ] ], "metrics_loc": "column", "metrics_type": "['similarity', 'similarity', 'similarity', 'similarity', 'similarity', 'similarity', 'similarity', 'similarity', 'similarity', 'similarity', 'similarity', 'similarity']", "target_entity": "['VGGNet']", "table_html": "
Arch. || AlexNet || Agg. || Mean || Type/Eval || SL Arch. || AlexNet || Agg. || Mean || Type/Eval || MEN Arch. || AlexNet || Agg. || Max || Type/Eval || SL Arch. || AlexNet || Agg. || Max || Type/Eval || MEN Arch. || GoogLeNet || Agg. || Mean || Type/Eval || SL Arch. || GoogLeNet || Agg. || Mean || Type/Eval || MEN Arch. || GoogLeNet || Agg. || Max || Type/Eval || SL Arch. || GoogLeNet || Agg. || Max || Type/Eval || MEN Arch. || VGGNet || Agg. || Mean || Type/Eval || SL Arch. || VGGNet || Agg. || Mean || Type/Eval || MEN Arch. || VGGNet || Agg. || Max || Type/Eval || SL Arch. || VGGNet || Agg. || Max || Type/Eval || MEN
Source || Wikipedia || Text 0.310 0.682 0.310 0.682 0.310 0.682 0.310 0.682 0.310 0.682 0.310 0.682
Source || Google || Visual 0.340 0.503 0.334 0.513 0.358 0.495 0.367 0.501 0.342 0.512 0.332 0.494
Source || Google || MM 0.380 0.711 0.370 0.719 0.379 0.711 0.365 0.716 0.380 0.714 0.365 0.716
Source || Bing || Visual 0.325 0.567 0.316 0.554 0.310 0.526 0.303 0.520 0.304 0.551 0.289 0.507
Source || Bing || MM 0.373 0.727 0.360 0.725 0.364 0.723 0.350 0.724 0.361 0.727 0.349 0.719
Source || Flickr || Visual 0.234 0.483 0.224 0.441 0.238 0.407 0.236 0.385 0.243 0.460 0.226 0.385
Source || Flickr || MM 0.350 0.715 0.343 0.711 0.347 0.689 0.344 0.703 0.354 0.702 0.339 0.696
Source || ImageNet || Visual 0.313 0.561 0.313 0.561 0.341 0.540 0.411 0.603 0.404 0.584 0.401 0.578
Source || ImageNet || MM 0.362 0.713 0.362 0.713 0.373 0.719 0.401 0.731 0.427 0.727 0.412 0.723
Source || ESPGame || Visual 0.018 0.448 0.026 0.376 0.063 0.487 0.050 0.434 0.125 0.506 0.106 0.451
Source || ESPGame || MM 0.208 0.686 0.187 0.672 0.243 0.700 0.246 0.696 0.269 0.708 0.260 0.698
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D16-1043", "page_no": 5, "dir": "emnlp2016", "description": "5.1 Maximum coverage comparison. Table 4 shows the results on the maximally covered datasets. This means we cannot directly compare between data sources, because they have different coverage, but we can look at absolute performance and compare network architectures. The first row reports results for the text-based linguistic representations that were obtained from Wikipedia (repeated across columns for convenience). For each of the three architectures, we evaluate on SimLex (SL) and MEN, using either the mean (Mean) or elementwise maximum (Max) method for aggregating image representations into visual ones (see Section 2). For each data source, we report results for the visual representations, as well as for the multi-modal representations that fuse the visual and textual ones together. Performance across architectures is remarkably stable: we have had to report results up to three decimal points to show the difference in performance in some cases. For each of the network architectures, we see a marked improvement of multi-modal representations over uni-modal linguistic representations. In many cases, we also see visual representations outperforming linguistic ones, especially on SimLex. This is interesting, because e.g. Google and Bing have full coverage over the datasets, so their visual representations include highly abstract words, which does not appear to have an adverse impact on the method\u2019s performance. For the ESP Game dataset (on which performance is quite low) and ImageNet, we observe an increase in performance as we move to the right in the table. Interestingly, VGGNet on ImageNet scores very highly, which seems to indicate that VGGNet is somehow more \u201cspecialized\u201d on ImageNet than the others. The difference between mean and max aggregation is relatively small, although the former seems to work better for SimLex while the latter does slightly better for MEN.", "class_sentence": "[2, 1, 2, 2, 2, 2, 2, 1, 1, 2, 1, 1, 1]", "sentences": "['5.1 Maximum coverage comparison.', 'Table 4 shows the results on the maximally covered datasets.', 'This means we cannot directly compare between data sources, because they have different coverage, but we can look at absolute performance and compare network architectures.', 'The first row reports results for the text-based linguistic representations that were obtained from Wikipedia (repeated across columns for convenience).', 'For each of the three architectures, we evaluate on SimLex (SL) and MEN, using either the mean (Mean) or elementwise maximum (Max) method for aggregating image representations into visual ones (see Section 2).', 'For each data source, we report results for the visual representations, as well as for the multi-modal representations that fuse the visual and textual ones together.', 'Performance across architectures is remarkably stable: we have had to report results up to three decimal points to show the difference in performance in some cases.', 'For each of the network architectures, we see a marked improvement of multi-modal representations over uni-modal linguistic representations.', 'In many cases, we also see visual representations outperforming linguistic ones, especially on SimLex.', 'This is interesting, because e.g. Google and Bing have full coverage over the datasets, so their visual representations include highly abstract words, which does not appear to have an adverse impact on the method\u2019s performance.', 'For the ESP Game dataset (on which performance is quite low) and ImageNet, we observe an increase in performance as we move to the right in the table.', 'Interestingly, VGGNet on ImageNet scores very highly, which seems to indicate that VGGNet is somehow more \u201cspecialized\u201d on ImageNet than the others.', 'The difference between mean and max aggregation is relatively small, although the former seems to work better for SimLex while the latter does slightly better for MEN.']", "header_mention": "[None, None, None, None, ['SL', 'MEN', 'Mean', 'Max'], None, None, None, ['Visual', 'MM', 'SL'], ['Google', 'Bing'], ['ESPGame', 'ImageNet'], ['VGGNet', 'ImageNet'], ['Max', 'Mean', 'MEN', 'SL']]", "valid": 1, "image_id": "D16-1043-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance on maximally covered datasets.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Arch. || AlexNet || Agg. || Mean || Type/Eval || SLArch. || AlexNet || Agg. || Mean || Type/Eval || MENArch. || AlexNet || Agg. || Max || Type/Eval || SLArch. || AlexNet || Agg. || Max || Type/Eval || MENArch. || GoogLeNet || Agg. || Mean || Type/Eval || SLArch. || GoogLeNet || Agg. || Mean || Type/Eval || MENArch. || GoogLeNet || Agg. || Max || Type/Eval || SLArch. || GoogLeNet || Agg. || Max || Type/Eval || MENArch. || VGGNet || Agg. || Mean || Type/Eval || SLArch. || VGGNet || Agg. || Mean || Type/Eval || MENArch. || VGGNet || Agg. || Max || Type/Eval || SLArch. || VGGNet || Agg. || Max || Type/Eval || MEN
Source || Wikipedia || Text0.3100.6820.3100.6820.3100.6820.3100.6820.3100.6820.3100.682
Source || Google || Visual0.3400.5030.3340.5130.3580.4950.3670.5010.3420.5120.3320.494
Source || Google || MM0.3800.7110.3700.7190.3790.7110.3650.7160.3800.7140.3650.716
Source || Bing || Visual0.3250.5670.3160.5540.3100.5260.3030.5200.3040.5510.2890.507
Source || Bing || MM0.3730.7270.3600.7250.3640.7230.3500.7240.3610.7270.3490.719
Source || Flickr || Visual0.2340.4830.2240.4410.2380.4070.2360.3850.2430.4600.2260.385
Source || Flickr || MM0.3500.7150.3430.7110.3470.6890.3440.7030.3540.7020.3390.696
Source || ImageNet || Visual0.3130.5610.3130.5610.3410.5400.4110.6030.4040.5840.4010.578
Source || ImageNet || MM0.3620.7130.3620.7130.3730.7190.4010.7310.4270.7270.4120.723
Source || ESPGame || Visual0.0180.4480.0260.3760.0630.4870.0500.4340.1250.5060.1060.451
Source || ESPGame || MM0.2080.6860.1870.6720.2430.7000.2460.6960.2690.7080.2600.698
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance on maximally covered datasets.}\n\\begin{tabular}{lrrrrrrrrrrrr}\n\\toprule\n & Arch. $\\parallel$ AlexNet $\\parallel$ Agg. $\\parallel$ Mean $\\parallel$ Type/Eval $\\parallel$ SL & Arch. $\\parallel$ AlexNet $\\parallel$ Agg. $\\parallel$ Mean $\\parallel$ Type/Eval $\\parallel$ MEN & Arch. $\\parallel$ AlexNet $\\parallel$ Agg. $\\parallel$ Max $\\parallel$ Type/Eval $\\parallel$ SL & Arch. $\\parallel$ AlexNet $\\parallel$ Agg. $\\parallel$ Max $\\parallel$ Type/Eval $\\parallel$ MEN & Arch. $\\parallel$ GoogLeNet $\\parallel$ Agg. $\\parallel$ Mean $\\parallel$ Type/Eval $\\parallel$ SL & Arch. $\\parallel$ GoogLeNet $\\parallel$ Agg. $\\parallel$ Mean $\\parallel$ Type/Eval $\\parallel$ MEN & Arch. $\\parallel$ GoogLeNet $\\parallel$ Agg. $\\parallel$ Max $\\parallel$ Type/Eval $\\parallel$ SL & Arch. $\\parallel$ GoogLeNet $\\parallel$ Agg. $\\parallel$ Max $\\parallel$ Type/Eval $\\parallel$ MEN & Arch. $\\parallel$ VGGNet $\\parallel$ Agg. $\\parallel$ Mean $\\parallel$ Type/Eval $\\parallel$ SL & Arch. $\\parallel$ VGGNet $\\parallel$ Agg. $\\parallel$ Mean $\\parallel$ Type/Eval $\\parallel$ MEN & Arch. $\\parallel$ VGGNet $\\parallel$ Agg. $\\parallel$ Max $\\parallel$ Type/Eval $\\parallel$ SL & Arch. $\\parallel$ VGGNet $\\parallel$ Agg. $\\parallel$ Max $\\parallel$ Type/Eval $\\parallel$ MEN \\\\\n\\midrule\nSource $\\parallel$ Wikipedia $\\parallel$ Text & 0.31 & 0.682 & 0.31 & 0.682 & 0.31 & 0.682 & 0.31 & 0.682 & 0.31 & 0.682 & 0.31 & 0.682 \\\\\nSource $\\parallel$ Google $\\parallel$ Visual & 0.34 & 0.503 & 0.334 & 0.513 & 0.358 & 0.495 & 0.367 & 0.501 & 0.342 & 0.512 & 0.332 & 0.494 \\\\\nSource $\\parallel$ Google $\\parallel$ MM & 0.38 & 0.711 & 0.37 & 0.719 & 0.379 & 0.711 & 0.365 & 0.716 & 0.38 & 0.714 & 0.365 & 0.716 \\\\\nSource $\\parallel$ Bing $\\parallel$ Visual & 0.325 & 0.567 & 0.316 & 0.554 & 0.31 & 0.526 & 0.303 & 0.52 & 0.304 & 0.551 & 0.289 & 0.507 \\\\\nSource $\\parallel$ Bing $\\parallel$ MM & 0.373 & 0.727 & 0.36 & 0.725 & 0.364 & 0.723 & 0.35 & 0.724 & 0.361 & 0.727 & 0.349 & 0.719 \\\\\nSource $\\parallel$ Flickr $\\parallel$ Visual & 0.234 & 0.483 & 0.224 & 0.441 & 0.238 & 0.407 & 0.236 & 0.385 & 0.243 & 0.46 & 0.226 & 0.385 \\\\\nSource $\\parallel$ Flickr $\\parallel$ MM & 0.35 & 0.715 & 0.343 & 0.711 & 0.347 & 0.689 & 0.344 & 0.703 & 0.354 & 0.702 & 0.339 & 0.696 \\\\\nSource $\\parallel$ ImageNet $\\parallel$ Visual & 0.313 & 0.561 & 0.313 & 0.561 & 0.341 & 0.54 & 0.411 & 0.603 & 0.404 & 0.584 & 0.401 & 0.578 \\\\\nSource $\\parallel$ ImageNet $\\parallel$ MM & 0.362 & 0.713 & 0.362 & 0.713 & 0.373 & 0.719 & 0.401 & 0.731 & 0.427 & 0.727 & 0.412 & 0.723 \\\\\nSource $\\parallel$ ESPGame $\\parallel$ Visual & 0.018 & 0.448 & 0.026 & 0.376 & 0.063 & 0.487 & 0.05 & 0.434 & 0.125 & 0.506 & 0.106 & 0.451 \\\\\nSource $\\parallel$ ESPGame $\\parallel$ MM & 0.208 & 0.686 & 0.187 & 0.672 & 0.243 & 0.7 & 0.246 & 0.696 & 0.269 & 0.708 & 0.26 & 0.698 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Performance on maximally covered datasets.\n
\n \n Arch. || AlexNet || Agg. || Mean || Type/Eval || SL\n \n Arch. || AlexNet || Agg. || Mean || Type/Eval || MEN\n \n Arch. || AlexNet || Agg. || Max || Type/Eval || SL\n \n Arch. || AlexNet || Agg. || Max || Type/Eval || MEN\n \n Arch. || GoogLeNet || Agg. || Mean || Type/Eval || SL\n \n Arch. || GoogLeNet || Agg. || Mean || Type/Eval || MEN\n \n Arch. || GoogLeNet || Agg. || Max || Type/Eval || SL\n \n Arch. || GoogLeNet || Agg. || Max || Type/Eval || MEN\n \n Arch. || VGGNet || Agg. || Mean || Type/Eval || SL\n \n Arch. || VGGNet || Agg. || Mean || Type/Eval || MEN\n \n Arch. || VGGNet || Agg. || Max || Type/Eval || SL\n \n Arch. || VGGNet || Agg. || Max || Type/Eval || MEN\n
\n Source || Wikipedia || Text\n \n 0.310\n \n 0.682\n \n 0.310\n \n 0.682\n \n 0.310\n \n 0.682\n \n 0.310\n \n 0.682\n \n 0.310\n \n 0.682\n \n 0.310\n \n 0.682\n
\n Source || Google || Visual\n \n 0.340\n \n 0.503\n \n 0.334\n \n 0.513\n \n 0.358\n \n 0.495\n \n 0.367\n \n 0.501\n \n 0.342\n \n 0.512\n \n 0.332\n \n 0.494\n
\n Source || Google || MM\n \n 0.380\n \n 0.711\n \n 0.370\n \n 0.719\n \n 0.379\n \n 0.711\n \n 0.365\n \n 0.716\n \n 0.380\n \n 0.714\n \n 0.365\n \n 0.716\n
\n Source || Bing || Visual\n \n 0.325\n \n 0.567\n \n 0.316\n \n 0.554\n \n 0.310\n \n 0.526\n \n 0.303\n \n 0.520\n \n 0.304\n \n 0.551\n \n 0.289\n \n 0.507\n
\n Source || Bing || MM\n \n 0.373\n \n 0.727\n \n 0.360\n \n 0.725\n \n 0.364\n \n 0.723\n \n 0.350\n \n 0.724\n \n 0.361\n \n 0.727\n \n 0.349\n \n 0.719\n
\n Source || Flickr || Visual\n \n 0.234\n \n 0.483\n \n 0.224\n \n 0.441\n \n 0.238\n \n 0.407\n \n 0.236\n \n 0.385\n \n 0.243\n \n 0.460\n \n 0.226\n \n 0.385\n
\n Source || Flickr || MM\n \n 0.350\n \n 0.715\n \n 0.343\n \n 0.711\n \n 0.347\n \n 0.689\n \n 0.344\n \n 0.703\n \n 0.354\n \n 0.702\n \n 0.339\n \n 0.696\n
\n Source || ImageNet || Visual\n \n 0.313\n \n 0.561\n \n 0.313\n \n 0.561\n \n 0.341\n \n 0.540\n \n 0.411\n \n 0.603\n \n 0.404\n \n 0.584\n \n 0.401\n \n 0.578\n
\n Source || ImageNet || MM\n \n 0.362\n \n 0.713\n \n 0.362\n \n 0.713\n \n 0.373\n \n 0.719\n \n 0.401\n \n 0.731\n \n 0.427\n \n 0.727\n \n 0.412\n \n 0.723\n
\n Source || ESPGame || Visual\n \n 0.018\n \n 0.448\n \n 0.026\n \n 0.376\n \n 0.063\n \n 0.487\n \n 0.050\n \n 0.434\n \n 0.125\n \n 0.506\n \n 0.106\n \n 0.451\n
\n Source || ESPGame || MM\n \n 0.208\n \n 0.686\n \n 0.187\n \n 0.672\n \n 0.243\n \n 0.700\n \n 0.246\n \n 0.696\n \n 0.269\n \n 0.708\n \n 0.260\n \n 0.698\n
\n" }, { "instance_id": "0cd584e7-f7cb-4df2-9dbc-6fa07a85454c", "table_id_paper": "D16-1044table_4", "caption": "Open-ended and multiple-choice (MC) results on VQA test set (trained on train+val set) compared with state-of-the-art: accuracy in %. See Sec. 4.4.", "row_header_level": 1, "row_headers": [ [ "MCB" ], [ "MCB + Genome" ], [ "MCB + Att." ], [ "MCB + Att. + GloVe" ], [ "MCB + Att. + Genome" ], [ "MCB + Att. + GloVe + Genome" ], [ "Ensemble of 7 Att. models" ], [ "Naver Labs (challenge 2nd)" ], [ "HieCoAtt (Lu et al. 2016)" ], [ "DMN+ (Xiong et al. 2016)" ], [ "FDA (Ilievski et al. 2016)" ], [ "D-NMN (Andreas et al. 2016a)" ], [ "AMA (Wu et al. 2016)" ], [ "SAN (Yang et al. 2015)" ], [ "NMN (Andreas et al. 2016b)" ], [ "AYN (Malinowski et al. 2016)" ], [ "SMem (Xu and Saenko, 2016)" ], [ "VQA team (Antol et al. 2015)" ], [ "DPPnet (Noh et al. 2015)" ], [ "iBOWIMG (Zhou et al. 2015)" ] ], "column_header_level": 3, "column_headers": [ [ "Test-dev", "Open Ended", "Y/N" ], [ "Test-dev", "Open Ended", "No." ], [ "Test-dev", "Open Ended", "Other" ], [ "Test-dev", "Open Ended", "All" ], [ "Test-dev", "MC", "All" ], [ "Test-standard", "Open Ended", "Y/N" ], [ "Test-standard", "Open Ended", "No." ], [ "Test-standard", "Open Ended", "Other" ], [ "Test-standard", "Open Ended", "All" ], [ "Test-standard", "MC", "All" ] ], "contents": [ [ "81.2", "35.1", "49.3", "60.8", "65.4", "-", "-", "-", "-", "-" ], [ "81.7", "36.6", "51.5", "62.3", "66.4", "-", "-", "-", "-", "-" ], [ "82.2", "37.7", "54.8", "64.2", "68.6", "-", "-", "-", "-", "-" ], [ "82.5", "37.6", "55.6", "64.7", "69.1", "-", "-", "-", "-", "-" ], [ "81.7", "38.2", "57.0", "65.1", "69.5", "-", "-", "-", "-", "-" ], [ "82.3", "37.2", "57.4", "65.4", "69.9", "-", "-", "-", "-", "-" ], [ "83.4", "39.8", "58.5", "66.7", "70.2", "83.2", "39.5", "58.0", "66.5", "70.1" ], [ "83.5", "39.8", "54.8", "64.9", "69.4", "83.3", "38.7", "54.6", "64.8", "69.3" ], [ "79.7", "38.7", "51.7", "61.8", "65.8", "-", "-", "-", "62.1", "66.1" ], [ "80.5", "36.8", "48.3", "60.3", "-", "-", "-", "-", "60.4", "-" ], [ "81.1", "36.2", "45.8", "59.2", "-", "-", "-", "-", "59.5", "-" ], [ "81.1", "38.6", "45.5", "59.4", "-", "-", "-", "-", "59.4", "-" ], [ "81.0", "38.4", "45.2", "59.2", "-", "81.1", "37.1", "45.8", "59.4", "-" ], [ "79.3", "36.6", "46.1", "58.7", "-", "-", "-", "-", "58.9", "-" ], [ "81.2", "38.0", "44.0", "58.6", "-", "81.2", "37.7", "44.0", "58.7", "-" ], [ "78.4", "36.4", "46.3", "58.4", "-", "78.2", "36.3", "46.3", "58.4", "-" ], [ "80.9", "37.3", "43.1", "58.0", "-", "80.9", "37.5", "43.5", "58.2", "-" ], [ "80.5", "36.8", "43.1", "57.8", "62.7", "80.6", "36.5", "43.7", "58.2", "63.1" ], [ "80.7", "37.2", "41.7", "57.2", "-", "80.3", "36.9", "42.2", "57.4", "-" ], [ "76.5", "35.0", "42.6", "55.7", "-", "76.8", "35.0", "42.6", "55.9", "62.0" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy']", "target_entity": "['MCB', 'MCB + Genome', 'MCB + Att.', 'MCB + Att. + GloVe', 'MCB + Att. + Genome', 'MCB + Att. + GloVe + Genome', 'Ensemble of 7 Att. models']", "table_html": "
Test-dev || Open Ended || Y/N Test-dev || Open Ended || No. Test-dev || Open Ended || Other Test-dev || Open Ended || All Test-dev || MC || All Test-standard || Open Ended || Y/N Test-standard || Open Ended || No. Test-standard || Open Ended || Other Test-standard || Open Ended || All Test-standard || MC || All
MCB 81.2 35.1 49.3 60.8 65.4 - - - - -
MCB + Genome 81.7 36.6 51.5 62.3 66.4 - - - - -
MCB + Att. 82.2 37.7 54.8 64.2 68.6 - - - - -
MCB + Att. + GloVe 82.5 37.6 55.6 64.7 69.1 - - - - -
MCB + Att. + Genome 81.7 38.2 57.0 65.1 69.5 - - - - -
MCB + Att. + GloVe + Genome 82.3 37.2 57.4 65.4 69.9 - - - - -
Ensemble of 7 Att. models 83.4 39.8 58.5 66.7 70.2 83.2 39.5 58.0 66.5 70.1
Naver Labs (challenge 2nd) 83.5 39.8 54.8 64.9 69.4 83.3 38.7 54.6 64.8 69.3
HieCoAtt (Lu et al. 2016) 79.7 38.7 51.7 61.8 65.8 - - - 62.1 66.1
DMN+ (Xiong et al. 2016) 80.5 36.8 48.3 60.3 - - - - 60.4 -
FDA (Ilievski et al. 2016) 81.1 36.2 45.8 59.2 - - - - 59.5 -
D-NMN (Andreas et al. 2016a) 81.1 38.6 45.5 59.4 - - - - 59.4 -
AMA (Wu et al. 2016) 81.0 38.4 45.2 59.2 - 81.1 37.1 45.8 59.4 -
SAN (Yang et al. 2015) 79.3 36.6 46.1 58.7 - - - - 58.9 -
NMN (Andreas et al. 2016b) 81.2 38.0 44.0 58.6 - 81.2 37.7 44.0 58.7 -
AYN (Malinowski et al. 2016) 78.4 36.4 46.3 58.4 - 78.2 36.3 46.3 58.4 -
SMem (Xu and Saenko, 2016) 80.9 37.3 43.1 58.0 - 80.9 37.5 43.5 58.2 -
VQA team (Antol et al. 2015) 80.5 36.8 43.1 57.8 62.7 80.6 36.5 43.7 58.2 63.1
DPPnet (Noh et al. 2015) 80.7 37.2 41.7 57.2 - 80.3 36.9 42.2 57.4 -
iBOWIMG (Zhou et al. 2015) 76.5 35.0 42.6 55.7 - 76.8 35.0 42.6 55.9 62.0
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D16-1044", "page_no": 7, "dir": "emnlp2016", "description": "4.4 Comparison to State-of-the-Art. Table 4 compares our approach with the state-of-the-art on VQA test set. Our best single model uses MCB pooling with two attention maps. Additionally, we augment our training data with images and QA pairs from the Visual Genome dataset. We also concatenate the learned word embedding with pretrained GloVe vectors (Pennington et al. 2014). Each model in our ensemble of 7 models uses MCB with attention. Some of the models were trained with data from Visual Genome, and some were trained with two attention maps. This ensemble is 1.8 points above the next best approach on the VQA open-ended task and 0.8 points above the next best approach on the multiple-choice task (on Testdev). Even without ensembles, our \u201cMCB + Genome + Att. + GloVe\u201d model still outperforms the next best result by 0.5 points, with an accuracy of 65.4% versus 64.9% on the open-ended task (on Test-dev).", "class_sentence": "[2, 1, 2, 2, 2, 2, 2, 1, 1]", "sentences": "['4.4 Comparison to State-of-the-Art.', 'Table 4 compares our approach with the state-of-the-art on VQA test set.', 'Our best single model uses MCB pooling with two attention maps.', 'Additionally, we augment our training data with images and QA pairs from the Visual Genome dataset.', 'We also concatenate the learned word embedding with pretrained GloVe vectors (Pennington et al. 2014).', 'Each model in our ensemble of 7 models uses MCB with attention.', 'Some of the models were trained with data from Visual Genome, and some were trained with two attention maps.', 'This ensemble is 1.8 points above the next best approach on the VQA open-ended task and 0.8 points above the next best approach on the multiple-choice task (on Test-dev).', 'Even without ensembles, our \u201cMCB + Genome + Att. + GloVe\u201d model still outperforms the next best result by 0.5 points, with an accuracy of 65.4% versus 64.9% on the open-ended task (on Test-dev).']", "header_mention": "[None, None, None, None, None, ['Ensemble of 7 Att. models'], None, ['Ensemble of 7 Att. models', 'Test-dev'], ['MCB + Att. + GloVe + Genome', 'Naver Labs (challenge 2nd)', 'Test-dev']]", "valid": 1, "image_id": "D16-1044-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Open-ended and multiple-choice (MC) results on VQA test set (trained on train+val set) compared with state-of-the-art: accuracy in %. See Sec. 4.4.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Test-dev || Open Ended || Y/NTest-dev || Open Ended || No.Test-dev || Open Ended || OtherTest-dev || Open Ended || AllTest-dev || MC || AllTest-standard || Open Ended || Y/NTest-standard || Open Ended || No.Test-standard || Open Ended || OtherTest-standard || Open Ended || AllTest-standard || MC || All
MCB81.235.149.360.865.4-----
MCB + Genome81.736.651.562.366.4-----
MCB + Att.82.237.754.864.268.6-----
MCB + Att. + GloVe82.537.655.664.769.1-----
MCB + Att. + Genome81.738.257.065.169.5-----
MCB + Att. + GloVe + Genome82.337.257.465.469.9-----
Ensemble of 7 Att. models83.439.858.566.770.283.239.558.066.570.1
Naver Labs (challenge 2nd)83.539.854.864.969.483.338.754.664.869.3
HieCoAtt (Lu et al. 2016)79.738.751.761.865.8---62.166.1
DMN+ (Xiong et al. 2016)80.536.848.360.3----60.4-
FDA (Ilievski et al. 2016)81.136.245.859.2----59.5-
D-NMN (Andreas et al. 2016a)81.138.645.559.4----59.4-
AMA (Wu et al. 2016)81.038.445.259.2-81.137.145.859.4-
SAN (Yang et al. 2015)79.336.646.158.7----58.9-
NMN (Andreas et al. 2016b)81.238.044.058.6-81.237.744.058.7-
AYN (Malinowski et al. 2016)78.436.446.358.4-78.236.346.358.4-
SMem (Xu and Saenko, 2016)80.937.343.158.0-80.937.543.558.2-
VQA team (Antol et al. 2015)80.536.843.157.862.780.636.543.758.263.1
DPPnet (Noh et al. 2015)80.737.241.757.2-80.336.942.257.4-
iBOWIMG (Zhou et al. 2015)76.535.042.655.7-76.835.042.655.962.0
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Open-ended and multiple-choice (MC) results on VQA test set (trained on train+val set) compared with state-of-the-art: accuracy in \\%. See Sec. 4.4.}\n\\begin{tabular}{lrrrrllllll}\n\\toprule\n & Test-dev $\\parallel$ Open Ended $\\parallel$ Y/N & Test-dev $\\parallel$ Open Ended $\\parallel$ No. & Test-dev $\\parallel$ Open Ended $\\parallel$ Other & Test-dev $\\parallel$ Open Ended $\\parallel$ All & Test-dev $\\parallel$ MC $\\parallel$ All & Test-standard $\\parallel$ Open Ended $\\parallel$ Y/N & Test-standard $\\parallel$ Open Ended $\\parallel$ No. & Test-standard $\\parallel$ Open Ended $\\parallel$ Other & Test-standard $\\parallel$ Open Ended $\\parallel$ All & Test-standard $\\parallel$ MC $\\parallel$ All \\\\\n\\midrule\nMCB & 81.2 & 35.1 & 49.3 & 60.8 & 65.4 & - & - & - & - & - \\\\\nMCB + Genome & 81.7 & 36.6 & 51.5 & 62.3 & 66.4 & - & - & - & - & - \\\\\nMCB + Att. & 82.2 & 37.7 & 54.8 & 64.2 & 68.6 & - & - & - & - & - \\\\\nMCB + Att. + GloVe & 82.5 & 37.6 & 55.6 & 64.7 & 69.1 & - & - & - & - & - \\\\\nMCB + Att. + Genome & 81.7 & 38.2 & 57 & 65.1 & 69.5 & - & - & - & - & - \\\\\nMCB + Att. + GloVe + Genome & 82.3 & 37.2 & 57.4 & 65.4 & 69.9 & - & - & - & - & - \\\\\nEnsemble of 7 Att. models & 83.4 & 39.8 & 58.5 & 66.7 & 70.2 & 83.2 & 39.5 & 58.0 & 66.5 & 70.1 \\\\\nNaver Labs (challenge 2nd) & 83.5 & 39.8 & 54.8 & 64.9 & 69.4 & 83.3 & 38.7 & 54.6 & 64.8 & 69.3 \\\\\nHieCoAtt (Lu et al. 2016) & 79.7 & 38.7 & 51.7 & 61.8 & 65.8 & - & - & - & 62.1 & 66.1 \\\\\nDMN+ (Xiong et al. 2016) & 80.5 & 36.8 & 48.3 & 60.3 & - & - & - & - & 60.4 & - \\\\\nFDA (Ilievski et al. 2016) & 81.1 & 36.2 & 45.8 & 59.2 & - & - & - & - & 59.5 & - \\\\\nD-NMN (Andreas et al. 2016a) & 81.1 & 38.6 & 45.5 & 59.4 & - & - & - & - & 59.4 & - \\\\\nAMA (Wu et al. 2016) & 81 & 38.4 & 45.2 & 59.2 & - & 81.1 & 37.1 & 45.8 & 59.4 & - \\\\\nSAN (Yang et al. 2015) & 79.3 & 36.6 & 46.1 & 58.7 & - & - & - & - & 58.9 & - \\\\\nNMN (Andreas et al. 2016b) & 81.2 & 38 & 44 & 58.6 & - & 81.2 & 37.7 & 44.0 & 58.7 & - \\\\\nAYN (Malinowski et al. 2016) & 78.4 & 36.4 & 46.3 & 58.4 & - & 78.2 & 36.3 & 46.3 & 58.4 & - \\\\\nSMem (Xu and Saenko, 2016) & 80.9 & 37.3 & 43.1 & 58 & - & 80.9 & 37.5 & 43.5 & 58.2 & - \\\\\nVQA team (Antol et al. 2015) & 80.5 & 36.8 & 43.1 & 57.8 & 62.7 & 80.6 & 36.5 & 43.7 & 58.2 & 63.1 \\\\\nDPPnet (Noh et al. 2015) & 80.7 & 37.2 & 41.7 & 57.2 & - & 80.3 & 36.9 & 42.2 & 57.4 & - \\\\\niBOWIMG (Zhou et al. 2015) & 76.5 & 35 & 42.6 & 55.7 & - & 76.8 & 35.0 & 42.6 & 55.9 & 62.0 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Open-ended and multiple-choice (MC) results on VQA test set (trained on train+val set) compared with state-of-the-art: accuracy in %. See Sec. 4.4.\n
\n \n Test-dev || Open Ended || Y/N\n \n Test-dev || Open Ended || No.\n \n Test-dev || Open Ended || Other\n \n Test-dev || Open Ended || All\n \n Test-dev || MC || All\n \n Test-standard || Open Ended || Y/N\n \n Test-standard || Open Ended || No.\n \n Test-standard || Open Ended || Other\n \n Test-standard || Open Ended || All\n \n Test-standard || MC || All\n
\n MCB\n \n 81.2\n \n 35.1\n \n 49.3\n \n 60.8\n \n 65.4\n \n -\n \n -\n \n -\n \n -\n \n -\n
\n MCB + Genome\n \n 81.7\n \n 36.6\n \n 51.5\n \n 62.3\n \n 66.4\n \n -\n \n -\n \n -\n \n -\n \n -\n
\n MCB + Att.\n \n 82.2\n \n 37.7\n \n 54.8\n \n 64.2\n \n 68.6\n \n -\n \n -\n \n -\n \n -\n \n -\n
\n MCB + Att. + GloVe\n \n 82.5\n \n 37.6\n \n 55.6\n \n 64.7\n \n 69.1\n \n -\n \n -\n \n -\n \n -\n \n -\n
\n MCB + Att. + Genome\n \n 81.7\n \n 38.2\n \n 57.0\n \n 65.1\n \n 69.5\n \n -\n \n -\n \n -\n \n -\n \n -\n
\n MCB + Att. + GloVe + Genome\n \n 82.3\n \n 37.2\n \n 57.4\n \n 65.4\n \n 69.9\n \n -\n \n -\n \n -\n \n -\n \n -\n
\n Ensemble of 7 Att. models\n \n 83.4\n \n 39.8\n \n 58.5\n \n 66.7\n \n 70.2\n \n 83.2\n \n 39.5\n \n 58.0\n \n 66.5\n \n 70.1\n
\n Naver Labs (challenge 2nd)\n \n 83.5\n \n 39.8\n \n 54.8\n \n 64.9\n \n 69.4\n \n 83.3\n \n 38.7\n \n 54.6\n \n 64.8\n \n 69.3\n
\n HieCoAtt (Lu et al. 2016)\n \n 79.7\n \n 38.7\n \n 51.7\n \n 61.8\n \n 65.8\n \n -\n \n -\n \n -\n \n 62.1\n \n 66.1\n
\n DMN+ (Xiong et al. 2016)\n \n 80.5\n \n 36.8\n \n 48.3\n \n 60.3\n \n -\n \n -\n \n -\n \n -\n \n 60.4\n \n -\n
\n FDA (Ilievski et al. 2016)\n \n 81.1\n \n 36.2\n \n 45.8\n \n 59.2\n \n -\n \n -\n \n -\n \n -\n \n 59.5\n \n -\n
\n D-NMN (Andreas et al. 2016a)\n \n 81.1\n \n 38.6\n \n 45.5\n \n 59.4\n \n -\n \n -\n \n -\n \n -\n \n 59.4\n \n -\n
\n AMA (Wu et al. 2016)\n \n 81.0\n \n 38.4\n \n 45.2\n \n 59.2\n \n -\n \n 81.1\n \n 37.1\n \n 45.8\n \n 59.4\n \n -\n
\n SAN (Yang et al. 2015)\n \n 79.3\n \n 36.6\n \n 46.1\n \n 58.7\n \n -\n \n -\n \n -\n \n -\n \n 58.9\n \n -\n
\n NMN (Andreas et al. 2016b)\n \n 81.2\n \n 38.0\n \n 44.0\n \n 58.6\n \n -\n \n 81.2\n \n 37.7\n \n 44.0\n \n 58.7\n \n -\n
\n AYN (Malinowski et al. 2016)\n \n 78.4\n \n 36.4\n \n 46.3\n \n 58.4\n \n -\n \n 78.2\n \n 36.3\n \n 46.3\n \n 58.4\n \n -\n
\n SMem (Xu and Saenko, 2016)\n \n 80.9\n \n 37.3\n \n 43.1\n \n 58.0\n \n -\n \n 80.9\n \n 37.5\n \n 43.5\n \n 58.2\n \n -\n
\n VQA team (Antol et al. 2015)\n \n 80.5\n \n 36.8\n \n 43.1\n \n 57.8\n \n 62.7\n \n 80.6\n \n 36.5\n \n 43.7\n \n 58.2\n \n 63.1\n
\n DPPnet (Noh et al. 2015)\n \n 80.7\n \n 37.2\n \n 41.7\n \n 57.2\n \n -\n \n 80.3\n \n 36.9\n \n 42.2\n \n 57.4\n \n -\n
\n iBOWIMG (Zhou et al. 2015)\n \n 76.5\n \n 35.0\n \n 42.6\n \n 55.7\n \n -\n \n 76.8\n \n 35.0\n \n 42.6\n \n 55.9\n \n 62.0\n
\n" }, { "instance_id": "fa978f18-1d19-4997-ba2e-0f4a4c34ea28", "table_id_paper": "D16-1086table_1", "caption": "Corpus size (length in token) and system performance by genre. News* used gold trees and is not included in total.", "row_header_level": 2, "row_headers": [ [ "Genre", "News*" ], [ "Genre", "News" ], [ "Genre", "Wiki" ], [ "Genre", "Web" ], [ "Genre", "Total" ] ], "column_header_level": 1, "column_headers": [ [ "Sentences" ], [ "Length" ], [ "Yield" ], [ "Precision" ] ], "contents": [ [ "100", "19.3", "142", "78.9" ], [ "100", "19.3", "144", "70.8" ], [ "100", "21.4", "178", "61.8" ], [ "100", "19.2", "165", "49.1" ], [ "300", "20.0", "487", "60.2" ] ], "metrics_loc": "column", "metrics_type": "['Sentences', 'Length', 'Yield', 'Precision']", "target_entity": "['Genre']", "table_html": "
Sentences Length Yield Precision
Genre || News* 100 19.3 142 78.9
Genre || News 100 19.3 144 70.8
Genre || Wiki 100 21.4 178 61.8
Genre || Web 100 19.2 165 49.1
Genre || Total 300 20.0 487 60.2
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "D16-1086", "page_no": 5, "dir": "emnlp2016", "description": "Results. From the whole corpus of 300 sentences, PropsDE extracted 487 tuples, yielding on average 1.6 per sentence with 2.9 arguments. 60% of them were labeled as correct. Table 1 shows that most extractions are made from Wikipedia articles, whereas the highest precision can be observed for newswire text. According to our expectations, web pages are most challenging, presumably due to noisier language. These differences between the genres can also be seen in the precision-yield curve (Figure 2).", "class_sentence": "[2, 2, 2, 1, 1, 0]", "sentences": "['Results.', 'From the whole corpus of 300 sentences, PropsDE extracted 487 tuples, yielding on average 1.6 per sentence with 2.9 arguments.', '60% of them were labeled as correct.', 'Table 1 shows that most extractions are made from Wikipedia articles, whereas the highest precision can be observed for newswire text.', 'According to our expectations, web pages are most challenging, presumably due to noisier language.', 'These differences between the genres can also be seen in the precision-yield curve (Figure 2).']", "header_mention": "[None, None, None, ['Wiki', 'News'], ['Web'], None]", "valid": 1, "image_id": "D16-1086-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Corpus size (length in token) and system performance by genre. News* used gold trees and is not included in total.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n SentencesLengthYieldPrecision
Genre || News*10019.314278.9
Genre || News10019.314470.8
Genre || Wiki10021.417861.8
Genre || Web10019.216549.1
Genre || Total30020.048760.2
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Corpus size (length in token) and system performance by genre. News* used gold trees and is not included in total.}\n\\begin{tabular}{lrrrr}\n\\toprule\n & Sentences & Length & Yield & Precision \\\\\n\\midrule\nGenre $\\parallel$ News* & 100 & 19.3 & 142 & 78.9 \\\\\nGenre $\\parallel$ News & 100 & 19.3 & 144 & 70.8 \\\\\nGenre $\\parallel$ Wiki & 100 & 21.4 & 178 & 61.8 \\\\\nGenre $\\parallel$ Web & 100 & 19.2 & 165 & 49.1 \\\\\nGenre $\\parallel$ Total & 300 & 20 & 487 & 60.2 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Corpus size (length in token) and system performance by genre. News* used gold trees and is not included in total.\n
\n \n Sentences\n \n Length\n \n Yield\n \n Precision\n
\n Genre || News*\n \n 100\n \n 19.3\n \n 142\n \n 78.9\n
\n Genre || News\n \n 100\n \n 19.3\n \n 144\n \n 70.8\n
\n Genre || Wiki\n \n 100\n \n 21.4\n \n 178\n \n 61.8\n
\n Genre || Web\n \n 100\n \n 19.2\n \n 165\n \n 49.1\n
\n Genre || Total\n \n 300\n \n 20.0\n \n 487\n \n 60.2\n
\n" }, { "instance_id": "91567ac2-25fd-445c-ae50-5e62652faa79", "table_id_paper": "D16-1132table_3", "caption": "Results of instance-wise evaluation for anaphoric and cataphoric sets", "row_header_level": 4, "row_headers": [ [ "Set", "Anaphoric", "Method", "single-column CNN (w/ position vec.)" ], [ "Set", "Anaphoric", "Method", "MCNN (BASE)" ], [ "Set", "Anaphoric", "Method", "MCNN (BASE+SURFSEQ)" ], [ "Set", "Anaphoric", "Method", "MCNN (BASE+DEPTREE)" ], [ "Set", "Anaphoric", "Method", "MCNN (BASE+SURFSEQ+DEPTREE)" ], [ "Set", "Anaphoric", "Method", "MCNN (BASE+SURFSEQ+PREDCONTEXT)" ], [ "Set", "Anaphoric", "Method", "MCNN (BASE+DEPTREE+PREDCONTEXT)" ], [ "Set", "Anaphoric", "Method", "MCNN (Proposed)" ], [ "Set", "Cataphoric", "Method", "single-column CNN (w/ position vec.)" ], [ "Set", "Cataphoric", "Method", "MCNN (BASE)" ], [ "Set", "Cataphoric", "Method", "MCNN (BASE+SURFSEQ)" ], [ "Set", "Cataphoric", "Method", "MCNN (BASE+DEPTREE)" ], [ "Set", "Cataphoric", "Method", "MCNN (BASE+SURFSEQ+DEPTREE)" ], [ "Set", "Cataphoric", "Method", "MCNN (BASE+SURFSEQ+PREDCONTEXT)" ], [ "Set", "Cataphoric", "Method", "MCNN (BASE+DEPTREE+PREDCONTEXT)" ], [ "Set", "Cataphoric", "Method", "MCNN (Proposed)" ] ], "column_header_level": 1, "column_headers": [ [ "Recall" ], [ "Precision" ], [ "F-score" ], [ "Avg.P" ] ], "contents": [ [ "0.445", "0.525", "0.481", "0.341" ], [ "0.591", "0.33", "0.424", "0.367" ], [ "0.555", "0.566", "0.56", "0.565" ], [ "0.389", "0.615", "0.476", "0.518" ], [ "0.503", "0.66", "0.571", "0.599" ], [ "0.535", "0.611", "0.57", "0.581" ], [ "0.33", "0.699", "0.449", "0.528" ], [ "0.492", "0.673", "0.569", "0.602" ], [ "0.163", "0.293", "0.209", "0.163" ], [ "0.171", "0.13", "0.148", "0.099" ], [ "0.202", "0.417", "0.272", "0.257" ], [ "0.268", "0.438", "0.332", "0.329" ], [ "0.195", "0.525", "0.285", "0.33" ], [ "0.258", "0.406", "0.316", "0.276" ], [ "0.24", "0.488", "0.322", "0.341" ], [ "0.251", "0.522", "0.339", "0.337" ] ], "metrics_loc": "column", "metrics_type": "['Recall', 'Precision', 'F-score', 'Avg.P']", "target_entity": "['MCNN (Proposed)']", "table_html": "
Recall Precision F-score Avg.P
Set || Anaphoric || Method || single-column CNN (w/ position vec.) 0.445 0.525 0.481 0.341
Set || Anaphoric || Method || MCNN (BASE) 0.591 0.33 0.424 0.367
Set || Anaphoric || Method || MCNN (BASE+SURFSEQ) 0.555 0.566 0.56 0.565
Set || Anaphoric || Method || MCNN (BASE+DEPTREE) 0.389 0.615 0.476 0.518
Set || Anaphoric || Method || MCNN (BASE+SURFSEQ+DEPTREE) 0.503 0.66 0.571 0.599
Set || Anaphoric || Method || MCNN (BASE+SURFSEQ+PREDCONTEXT) 0.535 0.611 0.57 0.581
Set || Anaphoric || Method || MCNN (BASE+DEPTREE+PREDCONTEXT) 0.33 0.699 0.449 0.528
Set || Anaphoric || Method || MCNN (Proposed) 0.492 0.673 0.569 0.602
Set || Cataphoric || Method || single-column CNN (w/ position vec.) 0.163 0.293 0.209 0.163
Set || Cataphoric || Method || MCNN (BASE) 0.171 0.13 0.148 0.099
Set || Cataphoric || Method || MCNN (BASE+SURFSEQ) 0.202 0.417 0.272 0.257
Set || Cataphoric || Method || MCNN (BASE+DEPTREE) 0.268 0.438 0.332 0.329
Set || Cataphoric || Method || MCNN (BASE+SURFSEQ+DEPTREE) 0.195 0.525 0.285 0.33
Set || Cataphoric || Method || MCNN (BASE+SURFSEQ+PREDCONTEXT) 0.258 0.406 0.316 0.276
Set || Cataphoric || Method || MCNN (BASE+DEPTREE+PREDCONTEXT) 0.24 0.488 0.322 0.341
Set || Cataphoric || Method || MCNN (Proposed) 0.251 0.522 0.339 0.337
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D16-1132", "page_no": 9, "dir": "emnlp2016", "description": "The results in Table 3 show that our MCNN-based method achieved better average precision than the single-column CNN baseline except the method that uses only the BASE column set for the cataphoric case. The results also demonstrate that each column set consistently contributes to improving the average precision for both the anaphoric and cataphoric cases. However, Table 3 shows that the average precision for the cataphoric set remains low. As one future direction for further improvement, we need to explore clues for identifying cataphoric relations more accurately.", "class_sentence": "[1, 1, 1, 0]", "sentences": "['The results in Table 3 show that our MCNN-based method achieved better average precision than the single-column CNN baseline except the method that uses only the BASE column set for the cataphoric case.', 'The results also demonstrate that each column set consistently contributes to improving the average precision for both the anaphoric and cataphoric cases.', 'However, Table 3 shows that the average precision for the cataphoric set remains low.', 'As one future direction for further improvement, we need to explore clues for identifying cataphoric relations more accurately.']", "header_mention": "[['MCNN (Proposed)', 'single-column CNN (w/ position vec.)', 'MCNN (BASE)', 'Avg.P'], ['Avg.P', 'Anaphoric', 'Cataphoric'], ['Cataphoric', 'Avg.P'], None]", "valid": 1, "image_id": "D16-1132-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results of instance-wise evaluation for anaphoric and cataphoric sets

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n RecallPrecisionF-scoreAvg.P
Set || Anaphoric || Method || single-column CNN (w/ position vec.)0.4450.5250.4810.341
Set || Anaphoric || Method || MCNN (BASE)0.5910.330.4240.367
Set || Anaphoric || Method || MCNN (BASE+SURFSEQ)0.5550.5660.560.565
Set || Anaphoric || Method || MCNN (BASE+DEPTREE)0.3890.6150.4760.518
Set || Anaphoric || Method || MCNN (BASE+SURFSEQ+DEPTREE)0.5030.660.5710.599
Set || Anaphoric || Method || MCNN (BASE+SURFSEQ+PREDCONTEXT)0.5350.6110.570.581
Set || Anaphoric || Method || MCNN (BASE+DEPTREE+PREDCONTEXT)0.330.6990.4490.528
Set || Anaphoric || Method || MCNN (Proposed)0.4920.6730.5690.602
Set || Cataphoric || Method || single-column CNN (w/ position vec.)0.1630.2930.2090.163
Set || Cataphoric || Method || MCNN (BASE)0.1710.130.1480.099
Set || Cataphoric || Method || MCNN (BASE+SURFSEQ)0.2020.4170.2720.257
Set || Cataphoric || Method || MCNN (BASE+DEPTREE)0.2680.4380.3320.329
Set || Cataphoric || Method || MCNN (BASE+SURFSEQ+DEPTREE)0.1950.5250.2850.33
Set || Cataphoric || Method || MCNN (BASE+SURFSEQ+PREDCONTEXT)0.2580.4060.3160.276
Set || Cataphoric || Method || MCNN (BASE+DEPTREE+PREDCONTEXT)0.240.4880.3220.341
Set || Cataphoric || Method || MCNN (Proposed)0.2510.5220.3390.337
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results of instance-wise evaluation for anaphoric and cataphoric sets}\n\\begin{tabular}{lrrrr}\n\\toprule\n & Recall & Precision & F-score & Avg.P \\\\\n\\midrule\nSet $\\parallel$ Anaphoric $\\parallel$ Method $\\parallel$ single-column CNN (w/ position vec.) & 0.445 & 0.525 & 0.481 & 0.341 \\\\\nSet $\\parallel$ Anaphoric $\\parallel$ Method $\\parallel$ MCNN (BASE) & 0.591 & 0.33 & 0.424 & 0.367 \\\\\nSet $\\parallel$ Anaphoric $\\parallel$ Method $\\parallel$ MCNN (BASE+SURFSEQ) & 0.555 & 0.566 & 0.56 & 0.565 \\\\\nSet $\\parallel$ Anaphoric $\\parallel$ Method $\\parallel$ MCNN (BASE+DEPTREE) & 0.389 & 0.615 & 0.476 & 0.518 \\\\\nSet $\\parallel$ Anaphoric $\\parallel$ Method $\\parallel$ MCNN (BASE+SURFSEQ+DEPTREE) & 0.503 & 0.66 & 0.571 & 0.599 \\\\\nSet $\\parallel$ Anaphoric $\\parallel$ Method $\\parallel$ MCNN (BASE+SURFSEQ+PREDCONTEXT) & 0.535 & 0.611 & 0.57 & 0.581 \\\\\nSet $\\parallel$ Anaphoric $\\parallel$ Method $\\parallel$ MCNN (BASE+DEPTREE+PREDCONTEXT) & 0.33 & 0.699 & 0.449 & 0.528 \\\\\nSet $\\parallel$ Anaphoric $\\parallel$ Method $\\parallel$ MCNN (Proposed) & 0.492 & 0.673 & 0.569 & 0.602 \\\\\nSet $\\parallel$ Cataphoric $\\parallel$ Method $\\parallel$ single-column CNN (w/ position vec.) & 0.163 & 0.293 & 0.209 & 0.163 \\\\\nSet $\\parallel$ Cataphoric $\\parallel$ Method $\\parallel$ MCNN (BASE) & 0.171 & 0.13 & 0.148 & 0.099 \\\\\nSet $\\parallel$ Cataphoric $\\parallel$ Method $\\parallel$ MCNN (BASE+SURFSEQ) & 0.202 & 0.417 & 0.272 & 0.257 \\\\\nSet $\\parallel$ Cataphoric $\\parallel$ Method $\\parallel$ MCNN (BASE+DEPTREE) & 0.268 & 0.438 & 0.332 & 0.329 \\\\\nSet $\\parallel$ Cataphoric $\\parallel$ Method $\\parallel$ MCNN (BASE+SURFSEQ+DEPTREE) & 0.195 & 0.525 & 0.285 & 0.33 \\\\\nSet $\\parallel$ Cataphoric $\\parallel$ Method $\\parallel$ MCNN (BASE+SURFSEQ+PREDCONTEXT) & 0.258 & 0.406 & 0.316 & 0.276 \\\\\nSet $\\parallel$ Cataphoric $\\parallel$ Method $\\parallel$ MCNN (BASE+DEPTREE+PREDCONTEXT) & 0.24 & 0.488 & 0.322 & 0.341 \\\\\nSet $\\parallel$ Cataphoric $\\parallel$ Method $\\parallel$ MCNN (Proposed) & 0.251 & 0.522 & 0.339 & 0.337 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Results of instance-wise evaluation for anaphoric and cataphoric sets\n
\n \n Recall\n \n Precision\n \n F-score\n \n Avg.P\n
\n Set || Anaphoric || Method || single-column CNN (w/ position vec.)\n \n 0.445\n \n 0.525\n \n 0.481\n \n 0.341\n
\n Set || Anaphoric || Method || MCNN (BASE)\n \n 0.591\n \n 0.33\n \n 0.424\n \n 0.367\n
\n Set || Anaphoric || Method || MCNN (BASE+SURFSEQ)\n \n 0.555\n \n 0.566\n \n 0.56\n \n 0.565\n
\n Set || Anaphoric || Method || MCNN (BASE+DEPTREE)\n \n 0.389\n \n 0.615\n \n 0.476\n \n 0.518\n
\n Set || Anaphoric || Method || MCNN (BASE+SURFSEQ+DEPTREE)\n \n 0.503\n \n 0.66\n \n 0.571\n \n 0.599\n
\n Set || Anaphoric || Method || MCNN (BASE+SURFSEQ+PREDCONTEXT)\n \n 0.535\n \n 0.611\n \n 0.57\n \n 0.581\n
\n Set || Anaphoric || Method || MCNN (BASE+DEPTREE+PREDCONTEXT)\n \n 0.33\n \n 0.699\n \n 0.449\n \n 0.528\n
\n Set || Anaphoric || Method || MCNN (Proposed)\n \n 0.492\n \n 0.673\n \n 0.569\n \n 0.602\n
\n Set || Cataphoric || Method || single-column CNN (w/ position vec.)\n \n 0.163\n \n 0.293\n \n 0.209\n \n 0.163\n
\n Set || Cataphoric || Method || MCNN (BASE)\n \n 0.171\n \n 0.13\n \n 0.148\n \n 0.099\n
\n Set || Cataphoric || Method || MCNN (BASE+SURFSEQ)\n \n 0.202\n \n 0.417\n \n 0.272\n \n 0.257\n
\n Set || Cataphoric || Method || MCNN (BASE+DEPTREE)\n \n 0.268\n \n 0.438\n \n 0.332\n \n 0.329\n
\n Set || Cataphoric || Method || MCNN (BASE+SURFSEQ+DEPTREE)\n \n 0.195\n \n 0.525\n \n 0.285\n \n 0.33\n
\n Set || Cataphoric || Method || MCNN (BASE+SURFSEQ+PREDCONTEXT)\n \n 0.258\n \n 0.406\n \n 0.316\n \n 0.276\n
\n Set || Cataphoric || Method || MCNN (BASE+DEPTREE+PREDCONTEXT)\n \n 0.24\n \n 0.488\n \n 0.322\n \n 0.341\n
\n Set || Cataphoric || Method || MCNN (Proposed)\n \n 0.251\n \n 0.522\n \n 0.339\n \n 0.337\n
\n" }, { "instance_id": "153f7760-66f4-44c5-9fa3-fe5262ed91ce", "table_id_paper": "D16-1136table_3", "caption": "Bilingual Lexicon Induction performance from es, it, nl to en. Gouws and S\u00f8gaard (2015) + Panlex/Wikt is our reimplementation using Panlex/Wiktionary dictionary. All our models use Panlex as the dictionary. We reported the recall at 1 and 5. The best performance is bold.", "row_header_level": 2, "row_headers": [ [ "Model", "Gouws and S\u00f8gaard (2015) + Panlex" ], [ "Model", "Gouws and S\u00f8gaard (2015) + Wikt" ], [ "Model", "BilBOWA: Gouws et al. (2015)" ], [ "Model", "Vulic and Moens (2015)" ], [ "Model", "Our model (random selection)" ], [ "Model", "Our model (EM selection)" ], [ "Model", "Our model (EM selection) + Joint model" ], [ "Model", "Our model (EM selection) + combine embeddings (\u03b4 = 0.01)" ], [ "Model", "Our model (EM selection) + lemmatization" ] ], "column_header_level": 2, "column_headers": [ [ "es-en", "rec1" ], [ "es-en", "rec5" ], [ "it-en", "rec1" ], [ "it-en", "rec5" ], [ "nl-en", "rec1" ], [ "nl-en", "rec5" ], [ "Average", "rec1" ], [ "Average", "rec5" ] ], "contents": [ [ "37.6", "63.6", "26.6", "56.3", "49.8", "76.0", "38.0", "65.3" ], [ "61.6", "78.9", "62.6", "81.1", "65.6", "79.7", "63.3", "79.9" ], [ "51.6", "-", "55.7", "-", "57.5", "-", "54.9", "-" ], [ "68.9", "-", "68.3", "-", "39.2", "-", "58.8", "-" ], [ "41.1", "62.0", "57.4", "75.4", "34.3", "55.5", "44.3", "64.3" ], [ "67.3", "79.5", "66.8", "82.3", "64.7", "82.4", "66.3", "81.4" ], [ "68.0", "80.5", "70.5", "83.3", "68.8", "84.0", "69.1", "82.6" ], [ "74.7", "85.4", "80.8", "90.4", "79.1", "90.5", "78.2", "88.8" ], [ "74.9", "86.0", "81.3", "91.3", "79.8", "91.3", "78.7", "89.5" ] ], "metrics_loc": "column", "metrics_type": "['rec1', 'rec5', 'rec1', 'rec5', 'rec1', 'rec5', 'rec1', 'rec5']", "target_entity": "['Our model (random selection)', 'Our model (EM selection)', 'Our model (EM selection) + Joint model', 'Our model (EM selection) + combine embeddings (\u03b4 = 0.01)', 'Our model (EM selection) + lemmatization']", "table_html": "
es-en || rec1 es-en || rec5 it-en || rec1 it-en || rec5 nl-en || rec1 nl-en || rec5 Average || rec1 Average || rec5
Model || Gouws and S\u00f8gaard (2015) + Panlex 37.6 63.6 26.6 56.3 49.8 76.0 38.0 65.3
Model || Gouws and S\u00f8gaard (2015) + Wikt 61.6 78.9 62.6 81.1 65.6 79.7 63.3 79.9
Model || BilBOWA: Gouws et al. (2015) 51.6 - 55.7 - 57.5 - 54.9 -
Model || Vulic and Moens (2015) 68.9 - 68.3 - 39.2 - 58.8 -
Model || Our model (random selection) 41.1 62.0 57.4 75.4 34.3 55.5 44.3 64.3
Model || Our model (EM selection) 67.3 79.5 66.8 82.3 64.7 82.4 66.3 81.4
Model || Our model (EM selection) + Joint model 68.0 80.5 70.5 83.3 68.8 84.0 69.1 82.6
Model || Our model (EM selection) + combine embeddings (\u03b4 = 0.01) 74.7 85.4 80.8 90.4 79.1 90.5 78.2 88.8
Model || Our model (EM selection) + lemmatization 74.9 86.0 81.3 91.3 79.8 91.3 78.7 89.5
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D16-1136", "page_no": 7, "dir": "emnlp2016", "description": "Quantitative evaluation. Table 3 shows our results compared with prior work. We reimplement Gouws and S\u00f8gaard (2015) using Panlex and Wiktionary dictionaries. The result with Panlex is substantially worse than with Wiktionary. This confirms our hypothesis in \u00a72. That is the context might be corrupted if we just randomly replace the training data with the translation from noisy dictionary such as Panlex. Our model when randomly picking the translation is similar to Gouws and S\u00f8gaard (2015), using the Panlex dictionary. The biggest difference is that they replace the training data (both context and middle word) while we fix the context and only replace the middle word. For a high coverage yet noisy dictionary such as Panlex, our approach gives better average score. Comparing our two most basic models (EM selection and random selection), it is clear that the model using EM to select the translation outperforms random selection by a significant margin. Our joint model, as described in equation (3) which predicts both target word and the translation, further improves the performance, especially for nl-en. We use equation (5) to combine both context embeddings V and word embeddings U for all three language pairs. This modification during training substantially improves the performance. More importantly, all our improvements are consistent for all three language pairs and both evaluation metrics, showing the robustness of our models. Our combined model out-performed previous approaches by a large margin. Vulic and Moens (2015) used bilingual comparable data, but this might be hard to obtain for some language pairs. Their performance on nl-en is poor because their comparable data between en and nl is small. Besides, they also use POS tagger and lemmatizer to filter only Noun and reduce the morphology complexity during training. These tools might not be available for many languages. For a fairer comparison to their work, we also use the same Treetagger (Schmid, 1995) to lemmatize the output of our combined model before evaluation. Table 3 (+lemmatization) shows some improvements but minor. It demonstrates that our model is already good at disambiguating morphology. For example, the top 2 translations for es word lenguas in en are languages and language which correctly prefer the plural translation.", "class_sentence": "[2, 1, 2, 1, 2, 2, 1, 2, 1, 1, 1, 2, 1, 1, 1, 2, 1, 2, 2, 2, 1, 1, 2]", "sentences": "['Quantitative evaluation.', 'Table 3 shows our results compared with prior work.', 'We reimplement Gouws and S\u00f8gaard (2015) using Panlex and Wiktionary dictionaries.', 'The result with Panlex is substantially worse than with Wiktionary.', 'This confirms our hypothesis in \u00a72.', 'That is the context might be corrupted if we just randomly replace the training data with the translation from noisy dictionary such as Panlex.', 'Our model when randomly picking the translation is similar to Gouws and S\u00f8gaard (2015), using the Panlex dictionary.', 'The biggest difference is that they replace the training data (both context and middle word) while we fix the context and only replace the middle word.', 'For a high coverage yet noisy dictionary such as Panlex, our approach gives better average score.', 'Comparing our two most basic models (EM selection and random selection), it is clear that the model using EM to select the translation outperforms random selection by a significant margin.', 'Our joint model, as described in equation (3) which predicts both target word and the translation, further improves the performance, especially for nl-en.', 'We use equation (5) to combine both context embeddings V and word embeddings U for all three language pairs.', 'This modification during training substantially improves the performance.', 'More importantly, all our improvements are consistent for all three language pairs and both evaluation metrics, showing the robustness of our models.', 'Our combined model out-performed previous approaches by a large margin.', 'Vulic and Moens (2015) used bilingual comparable data, but this might be hard to obtain for some language pairs.', 'Their performance on nl-en is poor because their comparable data between en and nl is small.', 'Besides, they also use POS tagger and lemmatizer to filter only Noun and reduce the morphology complexity during training.', 'These tools might not be available for many languages.', 'For a fairer comparison to their work, we also use the same Treetagger (Schmid, 1995) to lemmatize the output of our combined model before evaluation.', 'Table 3 (+lemmatization) shows some improvements but minor.', 'It demonstrates that our model is already good at disambiguating morphology.', 'For example, the top 2 translations for es word lenguas in en are languages and language which correctly prefer the plural translation.']", "header_mention": "[None, ['Model'], ['Gouws and S\u00f8gaard (2015) + Panlex', 'Gouws and S\u00f8gaard (2015) + Wikt'], ['Gouws and S\u00f8gaard (2015) + Panlex', 'Gouws and S\u00f8gaard (2015) + Wikt'], None, None, ['Our model (random selection)', 'Gouws and S\u00f8gaard (2015) + Panlex'], ['Our model (random selection)', 'Gouws and S\u00f8gaard (2015) + Panlex'], ['Average'], ['Our model (random selection)', 'Our model (EM selection)'], ['Our model (EM selection) + Joint model', 'nl-en'], ['Our model (EM selection) + combine embeddings (\u03b4 = 0.01)'], ['Our model (EM selection) + combine embeddings (\u03b4 = 0.01)', 'rec1', 'rec5'], ['Our model (EM selection) + combine embeddings (\u03b4 = 0.01)', 'es-en', 'it-en', 'nl-en', 'rec1', 'rec5'], ['Our model (EM selection) + combine embeddings (\u03b4 = 0.01)'], ['Vulic and Moens (2015)'], ['Vulic and Moens (2015)', 'nl-en'], ['Vulic and Moens (2015)'], None, None, ['Our model (EM selection) + lemmatization'], ['Our model (random selection)', 'Our model (EM selection)'], ['es-en']]", "valid": 1, "image_id": "D16-1136-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Bilingual Lexicon Induction performance from es, it, nl to en. Gouws and S\u00f8gaard (2015) + Panlex/Wikt is our reimplementation using Panlex/Wiktionary dictionary. All our models use Panlex as the dictionary. We reported the recall at 1 and 5. The best performance is bold.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n es-en || rec1es-en || rec5it-en || rec1it-en || rec5nl-en || rec1nl-en || rec5Average || rec1Average || rec5
Model || Gouws and S\u00f8gaard (2015) + Panlex37.663.626.656.349.876.038.065.3
Model || Gouws and S\u00f8gaard (2015) + Wikt61.678.962.681.165.679.763.379.9
Model || BilBOWA: Gouws et al. (2015)51.6-55.7-57.5-54.9-
Model || Vulic and Moens (2015)68.9-68.3-39.2-58.8-
Model || Our model (random selection)41.162.057.475.434.355.544.364.3
Model || Our model (EM selection)67.379.566.882.364.782.466.381.4
Model || Our model (EM selection) + Joint model68.080.570.583.368.884.069.182.6
Model || Our model (EM selection) + combine embeddings (\u03b4 = 0.01)74.785.480.890.479.190.578.288.8
Model || Our model (EM selection) + lemmatization74.986.081.391.379.891.378.789.5
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Bilingual Lexicon Induction performance from es, it, nl to en. Gouws and S\u00f8gaard (2015) + Panlex/Wikt is our reimplementation using Panlex/Wiktionary dictionary. All our models use Panlex as the dictionary. We reported the recall at 1 and 5. The best performance is bold.}\n\\begin{tabular}{lrlrlrlrl}\n\\toprule\n & es-en $\\parallel$ rec1 & es-en $\\parallel$ rec5 & it-en $\\parallel$ rec1 & it-en $\\parallel$ rec5 & nl-en $\\parallel$ rec1 & nl-en $\\parallel$ rec5 & Average $\\parallel$ rec1 & Average $\\parallel$ rec5 \\\\\n\\midrule\nModel $\\parallel$ Gouws and S\u00f8gaard (2015) + Panlex & 37.6 & 63.6 & 26.6 & 56.3 & 49.8 & 76.0 & 38 & 65.3 \\\\\nModel $\\parallel$ Gouws and S\u00f8gaard (2015) + Wikt & 61.6 & 78.9 & 62.6 & 81.1 & 65.6 & 79.7 & 63.3 & 79.9 \\\\\nModel $\\parallel$ BilBOWA: Gouws et al. (2015) & 51.6 & - & 55.7 & - & 57.5 & - & 54.9 & - \\\\\nModel $\\parallel$ Vulic and Moens (2015) & 68.9 & - & 68.3 & - & 39.2 & - & 58.8 & - \\\\\nModel $\\parallel$ Our model (random selection) & 41.1 & 62.0 & 57.4 & 75.4 & 34.3 & 55.5 & 44.3 & 64.3 \\\\\nModel $\\parallel$ Our model (EM selection) & 67.3 & 79.5 & 66.8 & 82.3 & 64.7 & 82.4 & 66.3 & 81.4 \\\\\nModel $\\parallel$ Our model (EM selection) + Joint model & 68 & 80.5 & 70.5 & 83.3 & 68.8 & 84.0 & 69.1 & 82.6 \\\\\nModel $\\parallel$ Our model (EM selection) + combine embeddings ($\\delta$ $=$ 0.01) & 74.7 & 85.4 & 80.8 & 90.4 & 79.1 & 90.5 & 78.2 & 88.8 \\\\\nModel $\\parallel$ Our model (EM selection) + lemmatization & 74.9 & 86.0 & 81.3 & 91.3 & 79.8 & 91.3 & 78.7 & 89.5 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Bilingual Lexicon Induction performance from es, it, nl to en. Gouws and S\u00f8gaard (2015) + Panlex/Wikt is our reimplementation using Panlex/Wiktionary dictionary. All our models use Panlex as the dictionary. We reported the recall at 1 and 5. The best performance is bold.\n
\n \n es-en || rec1\n \n es-en || rec5\n \n it-en || rec1\n \n it-en || rec5\n \n nl-en || rec1\n \n nl-en || rec5\n \n Average || rec1\n \n Average || rec5\n
\n Model || Gouws and S\u00f8gaard (2015) + Panlex\n \n 37.6\n \n 63.6\n \n 26.6\n \n 56.3\n \n 49.8\n \n 76.0\n \n 38.0\n \n 65.3\n
\n Model || Gouws and S\u00f8gaard (2015) + Wikt\n \n 61.6\n \n 78.9\n \n 62.6\n \n 81.1\n \n 65.6\n \n 79.7\n \n 63.3\n \n 79.9\n
\n Model || BilBOWA: Gouws et al. (2015)\n \n 51.6\n \n -\n \n 55.7\n \n -\n \n 57.5\n \n -\n \n 54.9\n \n -\n
\n Model || Vulic and Moens (2015)\n \n 68.9\n \n -\n \n 68.3\n \n -\n \n 39.2\n \n -\n \n 58.8\n \n -\n
\n Model || Our model (random selection)\n \n 41.1\n \n 62.0\n \n 57.4\n \n 75.4\n \n 34.3\n \n 55.5\n \n 44.3\n \n 64.3\n
\n Model || Our model (EM selection)\n \n 67.3\n \n 79.5\n \n 66.8\n \n 82.3\n \n 64.7\n \n 82.4\n \n 66.3\n \n 81.4\n
\n Model || Our model (EM selection) + Joint model\n \n 68.0\n \n 80.5\n \n 70.5\n \n 83.3\n \n 68.8\n \n 84.0\n \n 69.1\n \n 82.6\n
\n Model || Our model (EM selection) + combine embeddings (\u03b4 = 0.01)\n \n 74.7\n \n 85.4\n \n 80.8\n \n 90.4\n \n 79.1\n \n 90.5\n \n 78.2\n \n 88.8\n
\n Model || Our model (EM selection) + lemmatization\n \n 74.9\n \n 86.0\n \n 81.3\n \n 91.3\n \n 79.8\n \n 91.3\n \n 78.7\n \n 89.5\n
\n" }, { "instance_id": "f51086a3-8c85-4196-ab89-b17b7543fad1", "table_id_paper": "D16-1168table_1", "caption": "METEOR results for different configuration of our model on STARdev, STARtest and CARTOON datasets.", "row_header_level": 3, "row_headers": [ [ "Model", "FULL", "-" ], [ "Model", "FULL", "-SEM" ], [ "Model", "FULL", "-SYN" ] ], "column_header_level": 1, "column_headers": [ [ "STARdev" ], [ "STARtest" ], [ "CARTOON" ] ], "contents": [ [ "31.82", "29.16", "32.08" ], [ "28.72", "25.55", "27.55" ], [ "31.92", "29.14", "32.04" ] ], "metrics_loc": "column", "metrics_type": "['METEOR', 'METEOR', 'METEOR']", "target_entity": "['Model']", "table_html": "
STARdev STARtest CARTOON
Model || FULL || - 31.82 29.16 32.08
Model || FULL || -SEM 28.72 25.55 27.55
Model || FULL || -SYN 31.92 29.14 32.04
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "D16-1168", "page_no": 7, "dir": "emnlp2016", "description": "Table 1 reports METEOR, we notice that removing the semantic coherence scores in -SEM hurts the performance compared to FULL, this confirms our claim that semantic compatibility is crucial for building coherent stories. On the other hand, -SYN performs similarly to FULL. Closer inspection of the -SYN system\u2019s output reveals a greater diversity in thematic elements as a result of the relaxed syntactic compatibility constraints. Hence it is more likely to have greater overlap with any of the reference rewrites, resulting in higher METEOR scores.", "class_sentence": "[1, 1, 1, 1]", "sentences": "['Table 1 reports METEOR, we notice that removing the semantic coherence scores in -SEM hurts the performance compared to FULL, this confirms our claim that semantic compatibility is crucial for building coherent stories.', 'On the other hand, -SYN performs similarly to FULL.', 'Closer inspection of the -SYN system\u2019s output reveals a greater diversity in thematic elements as a result of the relaxed syntactic compatibility constraints.', 'Hence it is more likely to have greater overlap with any of the reference rewrites, resulting in higher METEOR scores.']", "header_mention": "[['-SEM', 'FULL'], ['FULL', '-SYN'], ['-SYN'], ['-SYN']]", "valid": 1, "image_id": "D16-1168-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

METEOR results for different configuration of our model on STARdev, STARtest and CARTOON datasets.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n STARdevSTARtestCARTOON
Model || FULL || -31.8229.1632.08
Model || FULL || -SEM28.7225.5527.55
Model || FULL || -SYN31.9229.1432.04
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{METEOR results for different configuration of our model on STARdev, STARtest and CARTOON datasets.}\n\\begin{tabular}{lrrr}\n\\toprule\n & STARdev & STARtest & CARTOON \\\\\n\\midrule\nModel $\\parallel$ FULL $\\parallel$ - & 31.82 & 29.16 & 32.08 \\\\\nModel $\\parallel$ FULL $\\parallel$ -SEM & 28.72 & 25.55 & 27.55 \\\\\nModel $\\parallel$ FULL $\\parallel$ -SYN & 31.92 & 29.14 & 32.04 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: METEOR results for different configuration of our model on STARdev, STARtest and CARTOON datasets.\n
\n \n STARdev\n \n STARtest\n \n CARTOON\n
\n Model || FULL || -\n \n 31.82\n \n 29.16\n \n 32.08\n
\n Model || FULL || -SEM\n \n 28.72\n \n 25.55\n \n 27.55\n
\n Model || FULL || -SYN\n \n 31.92\n \n 29.14\n \n 32.04\n
\n" }, { "instance_id": "750881d8-a7bd-493a-8849-d8b35f6a7f20", "table_id_paper": "D16-1188table_2", "caption": "Accuracy results on the test sets. Bold font marks the best performance for a dataset. * indicates statistical significance of improvement over lasso at p < 0.05 using micro sign test for one of our models LSI, GoW and word2vec (underlined).", "row_header_level": 3, "row_headers": [ [ "20NG", "dataset", "science" ], [ "20NG", "dataset", "sports" ], [ "20NG", "dataset", "religion" ], [ "20NG", "dataset", "computer" ], [ "Sentiment", "dataset", "vote" ], [ "Sentiment", "dataset", "movie" ], [ "Sentiment", "dataset", "books" ], [ "Sentiment", "dataset", "dvd" ], [ "Sentiment", "dataset", "electr." ], [ "Sentiment", "dataset", "kitch." ] ], "column_header_level": 2, "column_headers": [ [ "no reg.", "-" ], [ "lasso", "-" ], [ "ridge", "-" ], [ "elastic", "-" ], [ "group lasso", "LDA" ], [ "group lasso", "LSI" ], [ "group lasso", "sentence" ], [ "group lasso", "GoW" ], [ "group lasso", "word2vec" ] ], "contents": [ [ "0.946", "0.916", "0.954", "0.954", "0.968", "0.968*", "0.942", "0.967", "0.968*" ], [ "0.908", "0.907", "0.925", "0.92", "0.959", "0.964*", "0.966", "0.959*", "0.946*" ], [ "0.894", "0.876", "0.895", "0.89", "0.918", "0.907*", "0.934", "0.911*", "0.916*" ], [ "0.846", "0.843", "0.869", "0.856", "0.891", "0.885*", "0.904", "0.885*", "0.911*" ], [ "0.606", "0.643", "0.616", "0.622", "0.658", "0.653", "0.656", "0.640", "0.651" ], [ "0.865", "0.86", "0.87", "0.875", "0.900", "0.895", "0.895", "0.895", "0.890" ], [ "0.75", "0.77", "0.76", "0.78", "0.790", "0.795", "0.785", "0.790", "0.800" ], [ "0.765", "0.735", "0.77", "0.76", "0.800", "0.805*", "0.785", "0.795*", "0.795*" ], [ "0.79", "0.8", "0.8", "0.825", "0.800", "0.815", "0.805", "0.820", "0.815" ], [ "0.76", "0.8", "0.775", "0.8", "0.845", "0.860*", "0.855", "0.840", "0.855*" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy']", "target_entity": "['group lasso']", "table_html": "
no reg. || - lasso || - ridge || - elastic || - group lasso || LDA group lasso || LSI group lasso || sentence group lasso || GoW group lasso || word2vec
20NG || dataset || science 0.946 0.916 0.954 0.954 0.968 0.968* 0.942 0.967 0.968*
20NG || dataset || sports 0.908 0.907 0.925 0.92 0.959 0.964* 0.966 0.959* 0.946*
20NG || dataset || religion 0.894 0.876 0.895 0.89 0.918 0.907* 0.934 0.911* 0.916*
20NG || dataset || computer 0.846 0.843 0.869 0.856 0.891 0.885* 0.904 0.885* 0.911*
Sentiment || dataset || vote 0.606 0.643 0.616 0.622 0.658 0.653 0.656 0.640 0.651
Sentiment || dataset || movie 0.865 0.86 0.87 0.875 0.900 0.895 0.895 0.895 0.890
Sentiment || dataset || books 0.75 0.77 0.76 0.78 0.790 0.795 0.785 0.790 0.800
Sentiment || dataset || dvd 0.765 0.735 0.77 0.76 0.800 0.805* 0.785 0.795* 0.795*
Sentiment || dataset || electr. 0.79 0.8 0.8 0.825 0.800 0.815 0.805 0.820 0.815
Sentiment || dataset || kitch. 0.76 0.8 0.775 0.8 0.845 0.860* 0.855 0.840 0.855*
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D16-1188", "page_no": 7, "dir": "emnlp2016", "description": "In Table 2 we report the results of our experiments on the aforementioned datasets, and we distinguish our proposed regularizers LSI, GoW, word2vec with underlining. Our results are inline and confirm that of (Yogatama and Smith, 2014a) showing the advantages of using structured regularizers in the text categorization task. The group based regularizers perform systematically better than the baseline ones.", "class_sentence": "[1, 1, 1]", "sentences": "['In Table 2 we report the results of our experiments on the aforementioned datasets, and we distinguish our proposed regularizers LSI, GoW, word2vec with underlining.', 'Our results are inline and confirm that of (Yogatama and Smith, 2014a) showing the advantages of using structured regularizers in the text categorization task.', 'The group based regularizers perform systematically better than the baseline ones.']", "header_mention": "[['dataset', 'LSI', 'GoW', 'word2vec'], None, ['group lasso']]", "valid": 1, "image_id": "D16-1188-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Accuracy results on the test sets. Bold font marks the best performance for a dataset. * indicates statistical significance of improvement over lasso at p < 0.05 using micro sign test for one of our models LSI, GoW and word2vec (underlined).

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n no reg. || -lasso || -ridge || -elastic || -group lasso || LDAgroup lasso || LSIgroup lasso || sentencegroup lasso || GoWgroup lasso || word2vec
20NG || dataset || science0.9460.9160.9540.9540.9680.968*0.9420.9670.968*
20NG || dataset || sports0.9080.9070.9250.920.9590.964*0.9660.959*0.946*
20NG || dataset || religion0.8940.8760.8950.890.9180.907*0.9340.911*0.916*
20NG || dataset || computer0.8460.8430.8690.8560.8910.885*0.9040.885*0.911*
Sentiment || dataset || vote0.6060.6430.6160.6220.6580.6530.6560.6400.651
Sentiment || dataset || movie0.8650.860.870.8750.9000.8950.8950.8950.890
Sentiment || dataset || books0.750.770.760.780.7900.7950.7850.7900.800
Sentiment || dataset || dvd0.7650.7350.770.760.8000.805*0.7850.795*0.795*
Sentiment || dataset || electr.0.790.80.80.8250.8000.8150.8050.8200.815
Sentiment || dataset || kitch.0.760.80.7750.80.8450.860*0.8550.8400.855*
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Accuracy results on the test sets. Bold font marks the best performance for a dataset. * indicates statistical significance of improvement over lasso at p $<$ 0.05 using micro sign test for one of our models LSI, GoW and word2vec (underlined).}\n\\begin{tabular}{lrrrrrlrll}\n\\toprule\n & no reg. $\\parallel$ - & lasso $\\parallel$ - & ridge $\\parallel$ - & elastic $\\parallel$ - & group lasso $\\parallel$ LDA & group lasso $\\parallel$ LSI & group lasso $\\parallel$ sentence & group lasso $\\parallel$ GoW & group lasso $\\parallel$ word2vec \\\\\n\\midrule\n20NG $\\parallel$ dataset $\\parallel$ science & 0.946 & 0.916 & 0.954 & 0.954 & 0.968 & 0.968* & 0.942 & 0.967 & 0.968* \\\\\n20NG $\\parallel$ dataset $\\parallel$ sports & 0.908 & 0.907 & 0.925 & 0.92 & 0.959 & 0.964* & 0.966 & 0.959* & 0.946* \\\\\n20NG $\\parallel$ dataset $\\parallel$ religion & 0.894 & 0.876 & 0.895 & 0.89 & 0.918 & 0.907* & 0.934 & 0.911* & 0.916* \\\\\n20NG $\\parallel$ dataset $\\parallel$ computer & 0.846 & 0.843 & 0.869 & 0.856 & 0.891 & 0.885* & 0.904 & 0.885* & 0.911* \\\\\nSentiment $\\parallel$ dataset $\\parallel$ vote & 0.606 & 0.643 & 0.616 & 0.622 & 0.658 & 0.653 & 0.656 & 0.640 & 0.651 \\\\\nSentiment $\\parallel$ dataset $\\parallel$ movie & 0.865 & 0.86 & 0.87 & 0.875 & 0.9 & 0.895 & 0.895 & 0.895 & 0.890 \\\\\nSentiment $\\parallel$ dataset $\\parallel$ books & 0.75 & 0.77 & 0.76 & 0.78 & 0.79 & 0.795 & 0.785 & 0.790 & 0.800 \\\\\nSentiment $\\parallel$ dataset $\\parallel$ dvd & 0.765 & 0.735 & 0.77 & 0.76 & 0.8 & 0.805* & 0.785 & 0.795* & 0.795* \\\\\nSentiment $\\parallel$ dataset $\\parallel$ electr. & 0.79 & 0.8 & 0.8 & 0.825 & 0.8 & 0.815 & 0.805 & 0.820 & 0.815 \\\\\nSentiment $\\parallel$ dataset $\\parallel$ kitch. & 0.76 & 0.8 & 0.775 & 0.8 & 0.845 & 0.860* & 0.855 & 0.840 & 0.855* \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Accuracy results on the test sets. Bold font marks the best performance for a dataset. * indicates statistical significance of improvement over lasso at p < 0.05 using micro sign test for one of our models LSI, GoW and word2vec (underlined).\n
\n \n no reg. || -\n \n lasso || -\n \n ridge || -\n \n elastic || -\n \n group lasso || LDA\n \n group lasso || LSI\n \n group lasso || sentence\n \n group lasso || GoW\n \n group lasso || word2vec\n
\n 20NG || dataset || science\n \n 0.946\n \n 0.916\n \n 0.954\n \n 0.954\n \n 0.968\n \n 0.968*\n \n 0.942\n \n 0.967\n \n 0.968*\n
\n 20NG || dataset || sports\n \n 0.908\n \n 0.907\n \n 0.925\n \n 0.92\n \n 0.959\n \n 0.964*\n \n 0.966\n \n 0.959*\n \n 0.946*\n
\n 20NG || dataset || religion\n \n 0.894\n \n 0.876\n \n 0.895\n \n 0.89\n \n 0.918\n \n 0.907*\n \n 0.934\n \n 0.911*\n \n 0.916*\n
\n 20NG || dataset || computer\n \n 0.846\n \n 0.843\n \n 0.869\n \n 0.856\n \n 0.891\n \n 0.885*\n \n 0.904\n \n 0.885*\n \n 0.911*\n
\n Sentiment || dataset || vote\n \n 0.606\n \n 0.643\n \n 0.616\n \n 0.622\n \n 0.658\n \n 0.653\n \n 0.656\n \n 0.640\n \n 0.651\n
\n Sentiment || dataset || movie\n \n 0.865\n \n 0.86\n \n 0.87\n \n 0.875\n \n 0.900\n \n 0.895\n \n 0.895\n \n 0.895\n \n 0.890\n
\n Sentiment || dataset || books\n \n 0.75\n \n 0.77\n \n 0.76\n \n 0.78\n \n 0.790\n \n 0.795\n \n 0.785\n \n 0.790\n \n 0.800\n
\n Sentiment || dataset || dvd\n \n 0.765\n \n 0.735\n \n 0.77\n \n 0.76\n \n 0.800\n \n 0.805*\n \n 0.785\n \n 0.795*\n \n 0.795*\n
\n Sentiment || dataset || electr.\n \n 0.79\n \n 0.8\n \n 0.8\n \n 0.825\n \n 0.800\n \n 0.815\n \n 0.805\n \n 0.820\n \n 0.815\n
\n Sentiment || dataset || kitch.\n \n 0.76\n \n 0.8\n \n 0.775\n \n 0.8\n \n 0.845\n \n 0.860*\n \n 0.855\n \n 0.840\n \n 0.855*\n
\n" }, { "instance_id": "467abd26-910b-4c85-9e5d-b83b9b64bc2a", "table_id_paper": "D16-1216table_1", "caption": "Accuracy Results on Wikipedia and Stack Exchange.", "row_header_level": 2, "row_headers": [ [ "Model", "Bag-of-Words" ], [ "Model", "Linguistic Features" ], [ "Model", "With Discovered Features" ], [ "Model", "CNN" ] ], "column_header_level": 1, "column_headers": [ [ "Wiki" ], [ "SE" ] ], "contents": [ [ "80.9%", "64.6%" ], [ "82.6%", "65.2%" ], [ "83.8%", "65.7%" ], [ "85.8%", "66.4%" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy']", "target_entity": "['CNN']", "table_html": "
Wiki SE
Model || Bag-of-Words 80.9% 64.6%
Model || Linguistic Features 82.6% 65.2%
Model || With Discovered Features 83.8% 65.7%
Model || CNN 85.8% 66.4%
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "D16-1216", "page_no": 3, "dir": "emnlp2016", "description": "Table 1 first presents our reproduced classification accuracy test results (two labels: positive or negative politeness) for the bag-of-words and linguistic features based models of Danescu-Niculescu-Mizil et al. (2013) (for our dataset splits) as well as the performance of our CNN model. As seen, without using any manually defined, theory-inspired linguistic features, the simple CNN model performs better than the feature-based methods. Next, we also show how the linguistic features baseline improves on adding our novelly discovered features (plus correcting some exising features), revealed via the analysis in Sec. 6. Thus, this reduces the gap in performance between the linguistic features baseline and the CNN, and in turn provides a quantitative reasoning for the success of the CNN model. More details in Sec. 6.", "class_sentence": "[1, 1, 2, 2, 0]", "sentences": "['Table 1 presents our reproduced classification accuracy test results (two labels: positive or negative politeness) for the bag-of-words and linguistic features based models of Danescu-Niculescu-Mizil et al. (2013) (for our dataset splits) as well as the performance of our CNN model.', 'As seen, without using any manually defined, theory-inspired linguistic features, the simple CNN model performs better than the feature-based methods.', 'Next, we also show how the linguistic features baseline improves on adding our novelly discovered features (plus correcting some exising features), revealed via the analysis in Sec. 6.', 'Thus, this reduces the gap in performance between the linguistic features baseline and the CNN, and in turn provides a quantitative reasoning for the success of the CNN model.', 'More details in Sec. 6.']", "header_mention": "[['Wiki', 'SE', 'Bag-of-Words', 'Linguistic Features', 'With Discovered Features', 'CNN'], ['CNN'], None, None, None]", "valid": 1, "image_id": "D16-1216-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Accuracy Results on Wikipedia and Stack Exchange.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n WikiSE
Model || Bag-of-Words80.9%64.6%
Model || Linguistic Features82.6%65.2%
Model || With Discovered Features83.8%65.7%
Model || CNN85.8%66.4%
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Accuracy Results on Wikipedia and Stack Exchange.}\n\\begin{tabular}{lll}\n\\toprule\n & Wiki & SE \\\\\n\\midrule\nModel $\\parallel$ Bag-of-Words & 80.9\\% & 64.6\\% \\\\\nModel $\\parallel$ Linguistic Features & 82.6\\% & 65.2\\% \\\\\nModel $\\parallel$ With Discovered Features & 83.8\\% & 65.7\\% \\\\\nModel $\\parallel$ CNN & 85.8\\% & 66.4\\% \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Accuracy Results on Wikipedia and Stack Exchange.\n
\n \n Wiki\n \n SE\n
\n Model || Bag-of-Words\n \n 80.9%\n \n 64.6%\n
\n Model || Linguistic Features\n \n 82.6%\n \n 65.2%\n
\n Model || With Discovered Features\n \n 83.8%\n \n 65.7%\n
\n Model || CNN\n \n 85.8%\n \n 66.4%\n
\n" }, { "instance_id": "c22ee3bc-dc80-4c9d-ace9-41e3a7d1c590", "table_id_paper": "D16-1220table_3", "caption": "Cross-validation performance on VUAMC. B is always significantly different from N (p < .001), and B \u222a N is always significantly different from both B and N (p < .001).", "row_header_level": 4, "row_headers": [ [ "Genre", "News", "Features", "B" ], [ "Genre", "News", "Features", "N" ], [ "Genre", "News", "Features", "B \u222a N" ], [ "Genre", "Academic", "Features", "B" ], [ "Genre", "Academic", "Features", "N" ], [ "Genre", "Academic", "Features", "B \u222a N" ], [ "Genre", "Conversation", "Features", "B" ], [ "Genre", "Conversation", "Features", "N" ], [ "Genre", "Conversation", "Features", "B \u222a N" ], [ "Genre", "Fiction", "Features", "B" ], [ "Genre", "Fiction", "Features", "N" ], [ "Genre", "Fiction", "Features", "B \u222a N" ] ], "column_header_level": 1, "column_headers": [ [ "C" ], [ "P" ], [ "R" ], [ "F" ] ], "contents": [ [ "1.0", "0.475", "0.742", "0.576" ], [ "1.0", "0.576", "0.479", "0.522" ], [ "1.0", "0.615", "0.539", "0.574" ], [ "0.6", "0.489", "0.733", "0.568" ], [ "0.6", "0.572", "0.494", "0.511" ], [ "1.0", "0.539", "0.648", "0.569" ], [ "0.6", "0.292", "0.799", "0.416" ], [ "0.6", "0.304", "0.626", "0.393" ], [ "1.0", "0.299", "0.731", "0.406" ], [ "0.6", "0.349", "0.695", "0.460" ], [ "0.6", "0.430", "0.418", "0.421" ], [ "0.6", "0.409", "0.551", "0.465" ] ], "metrics_loc": "column", "metrics_type": "['C', 'P', 'R', 'F']", "target_entity": "['Features']", "table_html": "
C P R F
Genre || News || Features || B 1.0 0.475 0.742 0.576
Genre || News || Features || N 1.0 0.576 0.479 0.522
Genre || News || Features || B \u222a N 1.0 0.615 0.539 0.574
Genre || Academic || Features || B 0.6 0.489 0.733 0.568
Genre || Academic || Features || N 0.6 0.572 0.494 0.511
Genre || Academic || Features || B \u222a N 1.0 0.539 0.648 0.569
Genre || Conversation || Features || B 0.6 0.292 0.799 0.416
Genre || Conversation || Features || N 0.6 0.304 0.626 0.393
Genre || Conversation || Features || B \u222a N 1.0 0.299 0.731 0.406
Genre || Fiction || Features || B 0.6 0.349 0.695 0.460
Genre || Fiction || Features || N 0.6 0.430 0.418 0.421
Genre || Fiction || Features || B \u222a N 0.6 0.409 0.551 0.465
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D16-1220", "page_no": 4, "dir": "emnlp2016", "description": "Finally, Table 3 shows the effect of the different feature sets on VUAMC used by Klebanov et al. (2014). We use the same 12-fold data split as Klebanov et al. (2014), and also in this case we perform a grid-search to optimize the meta-parameter C of the logistic regression classifier. The best value of C identified for each genre and feature set is shown in the column labeled C. On this data, N features alone are significantly outperformed by B (p < 0.01). On the other hand, for the genres \u201cacademic\u201d and \u201cfiction\u201d, combining N and B features improves classification performance over B, and the difference is always statistically significant. Besides, the addition of N always leads to more balanced models, by compensating for the relatively lower precision of B. Due to the lack of a separate test set, as in the original setup by Klebanov et al. (2014), and to the high dimensionality of B\u2019s lexicalized features, we cannot rule out over-fitting as an explanation for the relatively good performance of B on this benchmark. It should also be noted that the results reported in (Klebanov et al., 2014) are not the same, due to the mentioned differences in the implementation of the features and possibly other differences in the experimental setup (e.g., data filtering, pre-processing and meta-parameter optimization). In particular, our implementation of the B features performs better than reported by Klebanov et al. (2014) on all four genres, namely: 0.52 vs. 0.51 for \u201cnews\u201d, 0.51 vs. 0.28 for \u201cacademic\u201d, 0.39 vs. 0.28 for \u201cconversation\u201d and 0.42 vs. 0.33 for \u201cfiction\u201d.", "class_sentence": "[1, 2, 2, 1, 1, 1, 2, 2, 2]", "sentences": "['Finally, Table 3 shows the effect of the different feature sets on VUAMC used by Klebanov et al. (2014).', 'We use the same 12-fold data split as Klebanov et al. (2014), and also in this case we perform a grid-search to optimize the meta-parameter C of the logistic regression classifier.', 'The best value of C identified for each genre and feature set is shown in the column labeled C.', 'On this data, N features alone are significantly outperformed by B (p < 0.01).', 'On the other hand, for the genres \u201cacademic\u201d and \u201cfiction\u201d, combining N and B features improves classification performance over B, and the difference is always statistically significant.', 'Besides, the addition of N always leads to more balanced models, by compensating for the relatively lower precision of B.', 'Due to the lack of a separate test set, as in the original setup by Klebanov et al. (2014), and to the high dimensionality of B\u2019s lexicalized features, we cannot rule out over-fitting as an explanation for the relatively good performance of B on this benchmark.', 'It should also be noted that the results reported in (Klebanov et al., 2014) are not the same, due to the mentioned differences in the implementation of the features and possibly other differences in the experimental setup (e.g., data filtering, pre-processing and meta-parameter optimization).', 'In particular, our implementation of the B features performs better than reported by Klebanov et al. (2014) on all four genres, namely: 0.52 vs. 0.51 for \u201cnews\u201d, 0.51 vs. 0.28 for \u201cacademic\u201d, 0.39 vs. 0.28 for \u201cconversation\u201d and 0.42 vs. 0.33 for \u201cfiction\u201d.']", "header_mention": "[None, None, ['C', 'Genre', 'Features'], ['N', 'B'], ['Academic', 'Fiction', 'B \u222a N', 'B'], ['N', 'P'], None, None, None]", "valid": 1, "image_id": "D16-1220-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Cross-validation performance on VUAMC. B is always significantly different from N (p < .001), and B \u222a N is always significantly different from both B and N (p < .001).

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n CPRF
Genre || News || Features || B1.00.4750.7420.576
Genre || News || Features || N1.00.5760.4790.522
Genre || News || Features || B \u222a N1.00.6150.5390.574
Genre || Academic || Features || B0.60.4890.7330.568
Genre || Academic || Features || N0.60.5720.4940.511
Genre || Academic || Features || B \u222a N1.00.5390.6480.569
Genre || Conversation || Features || B0.60.2920.7990.416
Genre || Conversation || Features || N0.60.3040.6260.393
Genre || Conversation || Features || B \u222a N1.00.2990.7310.406
Genre || Fiction || Features || B0.60.3490.6950.460
Genre || Fiction || Features || N0.60.4300.4180.421
Genre || Fiction || Features || B \u222a N0.60.4090.5510.465
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Cross-validation performance on VUAMC. B is always significantly different from N (p $<$ .001), and B $\\cup$ N is always significantly different from both B and N (p $<$ .001).}\n\\begin{tabular}{lrrrr}\n\\toprule\n & C & P & R & F \\\\\n\\midrule\nGenre $\\parallel$ News $\\parallel$ Features $\\parallel$ B & 1 & 0.475 & 0.742 & 0.576 \\\\\nGenre $\\parallel$ News $\\parallel$ Features $\\parallel$ N & 1 & 0.576 & 0.479 & 0.522 \\\\\nGenre $\\parallel$ News $\\parallel$ Features $\\parallel$ B $\\cup$ N & 1 & 0.615 & 0.539 & 0.574 \\\\\nGenre $\\parallel$ Academic $\\parallel$ Features $\\parallel$ B & 0.6 & 0.489 & 0.733 & 0.568 \\\\\nGenre $\\parallel$ Academic $\\parallel$ Features $\\parallel$ N & 0.6 & 0.572 & 0.494 & 0.511 \\\\\nGenre $\\parallel$ Academic $\\parallel$ Features $\\parallel$ B $\\cup$ N & 1 & 0.539 & 0.648 & 0.569 \\\\\nGenre $\\parallel$ Conversation $\\parallel$ Features $\\parallel$ B & 0.6 & 0.292 & 0.799 & 0.416 \\\\\nGenre $\\parallel$ Conversation $\\parallel$ Features $\\parallel$ N & 0.6 & 0.304 & 0.626 & 0.393 \\\\\nGenre $\\parallel$ Conversation $\\parallel$ Features $\\parallel$ B $\\cup$ N & 1 & 0.299 & 0.731 & 0.406 \\\\\nGenre $\\parallel$ Fiction $\\parallel$ Features $\\parallel$ B & 0.6 & 0.349 & 0.695 & 0.46 \\\\\nGenre $\\parallel$ Fiction $\\parallel$ Features $\\parallel$ N & 0.6 & 0.43 & 0.418 & 0.421 \\\\\nGenre $\\parallel$ Fiction $\\parallel$ Features $\\parallel$ B $\\cup$ N & 0.6 & 0.409 & 0.551 & 0.465 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Cross-validation performance on VUAMC. B is always significantly different from N (p < .001), and B \u222a N is always significantly different from both B and N (p < .001).\n
\n \n C\n \n P\n \n R\n \n F\n
\n Genre || News || Features || B\n \n 1.0\n \n 0.475\n \n 0.742\n \n 0.576\n
\n Genre || News || Features || N\n \n 1.0\n \n 0.576\n \n 0.479\n \n 0.522\n
\n Genre || News || Features || B \u222a N\n \n 1.0\n \n 0.615\n \n 0.539\n \n 0.574\n
\n Genre || Academic || Features || B\n \n 0.6\n \n 0.489\n \n 0.733\n \n 0.568\n
\n Genre || Academic || Features || N\n \n 0.6\n \n 0.572\n \n 0.494\n \n 0.511\n
\n Genre || Academic || Features || B \u222a N\n \n 1.0\n \n 0.539\n \n 0.648\n \n 0.569\n
\n Genre || Conversation || Features || B\n \n 0.6\n \n 0.292\n \n 0.799\n \n 0.416\n
\n Genre || Conversation || Features || N\n \n 0.6\n \n 0.304\n \n 0.626\n \n 0.393\n
\n Genre || Conversation || Features || B \u222a N\n \n 1.0\n \n 0.299\n \n 0.731\n \n 0.406\n
\n Genre || Fiction || Features || B\n \n 0.6\n \n 0.349\n \n 0.695\n \n 0.460\n
\n Genre || Fiction || Features || N\n \n 0.6\n \n 0.430\n \n 0.418\n \n 0.421\n
\n Genre || Fiction || Features || B \u222a N\n \n 0.6\n \n 0.409\n \n 0.551\n \n 0.465\n
\n" }, { "instance_id": "07c66b4a-4aed-48f7-9f17-1cb93b7fe17c", "table_id_paper": "D16-1262table_5", "caption": "Comparison of various decoders using the same model from our full system (Global A\u2217). We report F1 with and without the backoff model, the percentage of sentences that the decoder can parse, and the time spent decoding relative to A\u2217.", "row_header_level": 2, "row_headers": [ [ "Decoder", "Global A*" ], [ "Decoder", "Best-first" ], [ "Decoder", "10-best reranking" ], [ "Decoder", "100-best reranking" ], [ "Decoder", "2-best beam search" ], [ "Decoder", "4-best beam search" ], [ "Decoder", "8-best beam search" ] ], "column_header_level": 2, "column_headers": [ [ "Dev F1", "- backoff" ], [ "Dev F1", "backoff" ], [ "Relative Time", "-" ] ], "contents": [ [ "88.4", "88.4 (99.8%)", "1X" ], [ "87.5", "2.8 (6.7%)", "293.4X" ], [ "87.9", "87.9 (99.7%)", "8.5X" ], [ "88.2", "88.0 (99.4%)", "72.3X" ], [ "88.2", "85.7 (94.0%)", "2.0X" ], [ "88.3", "88.1 (99.2%)", "6.7X" ], [ "88.2", "86.8 (98.1%)", "26.3X" ] ], "metrics_loc": "column", "metrics_type": "['Dev F1', 'Dev F1', 'Relative Time']", "target_entity": "['Global A*']", "table_html": "
Dev F1 || - backoff Dev F1 || backoff Relative Time || -
Decoder || Global A* 88.4 88.4 (99.8%) 1X
Decoder || Best-first 87.5 2.8 (6.7%) 293.4X
Decoder || 10-best reranking 87.9 87.9 (99.7%) 8.5X
Decoder || 100-best reranking 88.2 88.0 (99.4%) 72.3X
Decoder || 2-best beam search 88.2 85.7 (94.0%) 2.0X
Decoder || 4-best beam search 88.3 88.1 (99.2%) 6.7X
Decoder || 8-best beam search 88.2 86.8 (98.1%) 26.3X
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "D16-1262", "page_no": 9, "dir": "emnlp2016", "description": "Lastly, to show that our parser is both more accurate and efficient than other decoding methods, we decode our full model using best-first search, reranking, and beam search. Table 5 shows the F1 scores with and without the backoff model, the portion of the sentences that each decoder is able to parse, and the time spent decoding relative to the A* parser. In the best-first search comparison, we do not include the informative A* heuristic, and the parser completes very few parses before reaching computational limits\u00d1showing the importance of heuristics in large search spaces. In the reranking comparison, we obtain n-best lists from the backoff model and rerank each result with the full model. In the beam search comparison, we use the approach from Clark et al. (2015) which greedily finds the top-n parses for each span in a bottom-up manner. Results indicate that both approximate methods are less accurate and slower than A*.", "class_sentence": "[2, 1, 2, 2, 2, 1]", "sentences": "['Lastly, to show that our parser is both more accurate and efficient than other decoding methods, we decode our full model using best-first search, reranking, and beam search.', 'Table 5 shows the F1 scores with and without the backoff model, the portion of the sentences that each decoder is able to parse, and the time spent decoding relative to the A* parser.', 'In the best-first search comparison, we do not include the informative A* heuristic, and the parser completes very few parses before reaching computational limits\u00d1showing the importance of heuristics in large search spaces.', 'In the reranking comparison, we obtain n-best lists from the backoff model and rerank each result with the full model.', 'In the beam search comparison, we use the approach from Clark et al. (2015) which greedily finds the top-n parses for each span in a bottom-up manner.', 'Results indicate that both approximate methods are less accurate and slower than A*.']", "header_mention": "[None, ['Dev F1', '- backoff', 'backoff', 'Relative Time'], ['Best-first'], ['10-best reranking', '100-best reranking'], ['2-best beam search', '4-best beam search', '8-best beam search'], ['Best-first', '10-best reranking', '100-best reranking', '2-best beam search', '4-best beam search', '8-best beam search', 'Global A*']]", "valid": 1, "image_id": "D16-1262-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison of various decoders using the same model from our full system (Global A\u2217). We report F1 with and without the backoff model, the percentage of sentences that the decoder can parse, and the time spent decoding relative to A\u2217.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Dev F1 || - backoffDev F1 || backoffRelative Time || -
Decoder || Global A*88.488.4 (99.8%)1X
Decoder || Best-first87.52.8 (6.7%)293.4X
Decoder || 10-best reranking87.987.9 (99.7%)8.5X
Decoder || 100-best reranking88.288.0 (99.4%)72.3X
Decoder || 2-best beam search88.285.7 (94.0%)2.0X
Decoder || 4-best beam search88.388.1 (99.2%)6.7X
Decoder || 8-best beam search88.286.8 (98.1%)26.3X
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison of various decoders using the same model from our full system (Global A$\\ast$). We report F1 with and without the backoff model, the percentage of sentences that the decoder can parse, and the time spent decoding relative to A$\\ast$.}\n\\begin{tabular}{lrll}\n\\toprule\n & Dev F1 $\\parallel$ - backoff & Dev F1 $\\parallel$ backoff & Relative Time $\\parallel$ - \\\\\n\\midrule\nDecoder $\\parallel$ Global A* & 88.4 & 88.4 (99.8\\%) & 1X \\\\\nDecoder $\\parallel$ Best-first & 87.5 & 2.8 (6.7\\%) & 293.4X \\\\\nDecoder $\\parallel$ 10-best reranking & 87.9 & 87.9 (99.7\\%) & 8.5X \\\\\nDecoder $\\parallel$ 100-best reranking & 88.2 & 88.0 (99.4\\%) & 72.3X \\\\\nDecoder $\\parallel$ 2-best beam search & 88.2 & 85.7 (94.0\\%) & 2.0X \\\\\nDecoder $\\parallel$ 4-best beam search & 88.3 & 88.1 (99.2\\%) & 6.7X \\\\\nDecoder $\\parallel$ 8-best beam search & 88.2 & 86.8 (98.1\\%) & 26.3X \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Comparison of various decoders using the same model from our full system (Global A\u2217). We report F1 with and without the backoff model, the percentage of sentences that the decoder can parse, and the time spent decoding relative to A\u2217.\n
\n \n Dev F1 || - backoff\n \n Dev F1 || backoff\n \n Relative Time || -\n
\n Decoder || Global A*\n \n 88.4\n \n 88.4 (99.8%)\n \n 1X\n
\n Decoder || Best-first\n \n 87.5\n \n 2.8 (6.7%)\n \n 293.4X\n
\n Decoder || 10-best reranking\n \n 87.9\n \n 87.9 (99.7%)\n \n 8.5X\n
\n Decoder || 100-best reranking\n \n 88.2\n \n 88.0 (99.4%)\n \n 72.3X\n
\n Decoder || 2-best beam search\n \n 88.2\n \n 85.7 (94.0%)\n \n 2.0X\n
\n Decoder || 4-best beam search\n \n 88.3\n \n 88.1 (99.2%)\n \n 6.7X\n
\n Decoder || 8-best beam search\n \n 88.2\n \n 86.8 (98.1%)\n \n 26.3X\n
\n" }, { "instance_id": "1da2db25-a07e-4273-ab1a-59c38f170c74", "table_id_paper": "D17-1241table_3", "caption": "SPE: Prediction Results", "row_header_level": 2, "row_headers": [ [ "Methods", "Unigram" ], [ "Methods", "LIWC" ], [ "Methods", "SVD" ], [ "Methods", "UserLDA" ], [ "Methods", "PostLDA_Word" ], [ "Methods", "PostLDA_Doc" ], [ "Methods", "Post-D-DM" ], [ "Methods", "User-D-DM" ], [ "Methods", "Post-D-DBOW" ], [ "Methods", "User-D-DBOW" ] ], "column_header_level": 1, "column_headers": [ [ "Tobacco" ], [ "Alcohol" ], [ "Drug" ] ], "contents": [ [ "0.663", "0.672", "0.644" ], [ "0.731", "0.689", "0.758" ], [ "0.779", "0.724", "0.764" ], [ "0.641", "0.603", "0.599" ], [ "0.733", "0.617", "0.628" ], [ "0.768", "0.687", "0.721" ], [ "0.536", "0.622", "0.52" ], [ "0.775", "0.73", "0.767" ], [ "0.531", "0.606", "0.526" ], [ "0.802", "0.768", "0.819" ] ], "metrics_loc": "column", "metrics_type": "['AUC', 'AUC', 'AUC']", "target_entity": "['User-D-DBOW']", "table_html": "
Tobacco Alcohol Drug
Methods || Unigram 0.663 0.672 0.644
Methods || LIWC 0.731 0.689 0.758
Methods || SVD 0.779 0.724 0.764
Methods || UserLDA 0.641 0.603 0.599
Methods || PostLDA_Word 0.733 0.617 0.628
Methods || PostLDA_Doc 0.768 0.687 0.721
Methods || Post-D-DM 0.536 0.622 0.52
Methods || User-D-DM 0.775 0.73 0.767
Methods || Post-D-DBOW 0.531 0.606 0.526
Methods || User-D-DBOW 0.802 0.768 0.819
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D17-1241", "page_no": 4, "dir": "emnlp2017", "description": "We evaluate the performance of our models using 10-fold cross validation. The evaluation results shown in Table 3 are based on weighted ROC AUC of the best models. Among all the feature learning methods for Facebook status updates, User-D-DBOW performed the best. It significantly outperformed all the baseline systems that only rely on supervised training (p < 0.01 based on t-tests). It also significantly outperformed all the traditional feature learning methods such as LDA and SVD (p < 0.01 based on t-tests). Moreover, in terms of whether to treat all the posts by the same user as one big document or separate documents, LDA prefers one post one document (models with a \"post\" prefix) while all the document vector-based methods prefer one user one document (models with a User prefix). Moreover, to use post-level LDA to derive the SPE of a user, the document-based aggregation method (PostLDA_Doc) performed better than the wordbased method (PostLDA_Word).", "class_sentence": "[1, 1, 1, 1, 1, 1, 1]", "sentences": "['We evaluate the performance of our models using 10-fold cross validation.', 'The evaluation results shown in Table 3 are based on weighted ROC AUC of the best models.', 'Among all the feature learning methods for Facebook status updates, User-D-DBOW performed the best.', 'It significantly outperformed all the baseline systems that only rely on supervised training (p < 0.01 based on t-tests).', 'It also significantly outperformed all the traditional feature learning methods such as LDA and SVD (p < 0.01 based on t-tests).', 'Moreover, in terms of whether to treat all the posts by the same user as one big document or separate documents, LDA prefers one post one document (models with a \"post\" prefix) while all the document vector-based methods prefer one user one document (models with a User prefix).', 'Moreover, to use post-level LDA to derive the SPE of a user, the document-based aggregation method (PostLDA_Doc) performed better than the wordbased method (PostLDA_Word).']", "header_mention": "[None, None, ['User-D-DBOW', 'Tobacco', 'Alcohol', 'Drug'], ['Unigram', 'LIWC'], ['SVD', 'UserLDA', 'PostLDA_Word', 'PostLDA_Doc'], ['PostLDA_Doc'], ['PostLDA_Doc', 'PostLDA_Word']]", "valid": 1, "image_id": "D17-1241-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

SPE: Prediction Results

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n TobaccoAlcoholDrug
Methods || Unigram0.6630.6720.644
Methods || LIWC0.7310.6890.758
Methods || SVD0.7790.7240.764
Methods || UserLDA0.6410.6030.599
Methods || PostLDA_Word0.7330.6170.628
Methods || PostLDA_Doc0.7680.6870.721
Methods || Post-D-DM0.5360.6220.52
Methods || User-D-DM0.7750.730.767
Methods || Post-D-DBOW0.5310.6060.526
Methods || User-D-DBOW0.8020.7680.819
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{SPE: Prediction Results}\n\\begin{tabular}{lrrr}\n\\toprule\n & Tobacco & Alcohol & Drug \\\\\n\\midrule\nMethods $\\parallel$ Unigram & 0.663 & 0.672 & 0.644 \\\\\nMethods $\\parallel$ LIWC & 0.731 & 0.689 & 0.758 \\\\\nMethods $\\parallel$ SVD & 0.779 & 0.724 & 0.764 \\\\\nMethods $\\parallel$ UserLDA & 0.641 & 0.603 & 0.599 \\\\\nMethods $\\parallel$ PostLDA\\_Word & 0.733 & 0.617 & 0.628 \\\\\nMethods $\\parallel$ PostLDA\\_Doc & 0.768 & 0.687 & 0.721 \\\\\nMethods $\\parallel$ Post-D-DM & 0.536 & 0.622 & 0.52 \\\\\nMethods $\\parallel$ User-D-DM & 0.775 & 0.73 & 0.767 \\\\\nMethods $\\parallel$ Post-D-DBOW & 0.531 & 0.606 & 0.526 \\\\\nMethods $\\parallel$ User-D-DBOW & 0.802 & 0.768 & 0.819 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: SPE: Prediction Results\n
\n \n Tobacco\n \n Alcohol\n \n Drug\n
\n Methods || Unigram\n \n 0.663\n \n 0.672\n \n 0.644\n
\n Methods || LIWC\n \n 0.731\n \n 0.689\n \n 0.758\n
\n Methods || SVD\n \n 0.779\n \n 0.724\n \n 0.764\n
\n Methods || UserLDA\n \n 0.641\n \n 0.603\n \n 0.599\n
\n Methods || PostLDA_Word\n \n 0.733\n \n 0.617\n \n 0.628\n
\n Methods || PostLDA_Doc\n \n 0.768\n \n 0.687\n \n 0.721\n
\n Methods || Post-D-DM\n \n 0.536\n \n 0.622\n \n 0.52\n
\n Methods || User-D-DM\n \n 0.775\n \n 0.73\n \n 0.767\n
\n Methods || Post-D-DBOW\n \n 0.531\n \n 0.606\n \n 0.526\n
\n Methods || User-D-DBOW\n \n 0.802\n \n 0.768\n \n 0.819\n
\n" }, { "instance_id": "d77922bc-7c86-4f89-a0e8-a5cf18b0fd5f", "table_id_paper": "D17-1245table_7", "caption": "Results obtained by the best model on each category of the test set for the factual vs opinion argument classification task", "row_header_level": 2, "row_headers": [ [ "Category", "fact" ], [ "Category", "opinion" ], [ "Category", "avg/total" ] ], "column_header_level": 1, "column_headers": [ [ "P" ], [ "R" ], [ "F1" ], [ "#argument per category" ] ], "contents": [ [ "0.49", "0.5", "0.5", "138" ], [ "0.88", "0.87", "0.88", "575" ], [ "0.81", "0.79", "0.8", "713" ] ], "metrics_loc": "column", "metrics_type": "['P', 'R', 'F1', '#argument per category']", "target_entity": "['avg/total']", "table_html": "
P R F1 #argument per category
Category || fact 0.49 0.5 0.5 138
Category || opinion 0.88 0.87 0.88 575
Category || avg/total 0.81 0.79 0.8 713
", "table_name": "Table 7", "table_id": "table_7", "paper_id": "D17-1245", "page_no": 4, "dir": "emnlp2017", "description": "Table 7 reports on the results obtained by the best configuration, i.e. LR + All features, per each category.", "class_sentence": "[1]", "sentences": "['Table 7 reports on the results obtained by the best configuration, i.e. LR + All features, per each category.']", "header_mention": "[['avg/total', 'F1', '#argument per category']]", "valid": 1, "image_id": "D17-1245-Table7-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results obtained by the best model on each category of the test set for the factual vs opinion argument classification task

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n PRF1#argument per category
Category || fact0.490.50.5138
Category || opinion0.880.870.88575
Category || avg/total0.810.790.8713
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{6}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results obtained by the best model on each category of the test set for the factual vs opinion argument classification task}\n\\begin{tabular}{lrrrr}\n\\toprule\n & P & R & F1 & \\#argument per category \\\\\n\\midrule\nCategory $\\parallel$ fact & 0.49 & 0.5 & 0.5 & 138 \\\\\nCategory $\\parallel$ opinion & 0.88 & 0.87 & 0.88 & 575 \\\\\nCategory $\\parallel$ avg/total & 0.81 & 0.79 & 0.8 & 713 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 7: Results obtained by the best model on each category of the test set for the factual vs opinion argument classification task\n
\n \n P\n \n R\n \n F1\n \n #argument per category\n
\n Category || fact\n \n 0.49\n \n 0.5\n \n 0.5\n \n 138\n
\n Category || opinion\n \n 0.88\n \n 0.87\n \n 0.88\n \n 575\n
\n Category || avg/total\n \n 0.81\n \n 0.79\n \n 0.8\n \n 713\n
\n" }, { "instance_id": "0de9fa86-b00b-4b9f-864a-c63bbf966930", "table_id_paper": "D17-1254table_4", "caption": "Results for image-sentence ranking experiments on the COCO dataset. R@K denotes Recall@K (higher is better) and Med r is the median rank (lower is better). (\u2020) taken from Kiros et al. (2015). (\u2217) taken from Karpathy and Fei-Fei (2015). (\u2021) taken from Mao et al. (2015).", "row_header_level": 2, "row_headers": [ [ "Method", "uni-skip" ], [ "Method", "bi-skip" ], [ "Method", "combine-skip" ], [ "Our Results", "hierarchical model+emb." ], [ "Our Results", "composite model+emb." ], [ "Our Results", "combine+emb." ], [ "Task-dependent methods", "DVSA*" ], [ "Task-dependent methods", "m-RNN" ] ], "column_header_level": 1, "column_headers": [ [ "R@1" ], [ "Med r" ], [ "R@2" ], [ "Med r" ] ], "contents": [ [ "30.6", "3", "22.7", "4" ], [ "32.7", "3", "24.2", "4" ], [ "33.8", "3", "25.9", "4" ], [ "32.7", "3", "25.3", "4" ], [ "33.8", "3", "25.7", "4" ], [ "34.4", "3", "26.6", "4" ], [ "38.4", "1", "27.4", "3" ], [ "41", "2", "29", "3" ] ], "metrics_loc": "column", "metrics_type": "['R@1', 'Med r', 'R@2', 'Med r']", "target_entity": "['combine+emb.']", "table_html": "
R@1 Med r R@2 Med r
Method || uni-skip 30.6 3 22.7 4
Method || bi-skip 32.7 3 24.2 4
Method || combine-skip 33.8 3 25.9 4
Our Results || hierarchical model+emb. 32.7 3 25.3 4
Our Results || composite model+emb. 33.8 3 25.7 4
Our Results || combine+emb. 34.4 3 26.6 4
Task-dependent methods || DVSA* 38.4 1 27.4 3
Task-dependent methods || m-RNN 41 2 29 3
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D17-1254", "page_no": 8, "dir": "emnlp2017", "description": "Table 4 shows our results. Consistent with previous experiments, we empirically found that the encoder trained using the fixed word embedding performed better on this task, hence only results using this method are reported. As can be seen, we obtain the same median rank as in Kiros et al.(2015), indicating that our encoder is as competitive as the skip-thought vectors (Kiros et al., 2015). The performance gain between our encoder and the combine-skip model of Kiros et al.(2015) on the R@1 score is significant, which shows that the CNN encoder has more discriminative power on retrieving the most correct item than the skip-thought vector.", "class_sentence": "[1, 2, 1, 1]", "sentences": "['Table 4 shows our results.', 'Consistent with previous experiments, we empirically found that the encoder trained using the fixed word embedding performed better on this task, hence only results using this method are reported.', 'As can be seen, we obtain the same median rank as in Kiros et al.(2015), indicating that our encoder is as competitive as the skip-thought vectors (Kiros et al., 2015).', 'The performance gain between our encoder and the combine-skip model of Kiros et al.(2015) on the R@1 score is significant, which shows that the CNN encoder has more discriminative power on retrieving the most correct item than the skip-thought vector.']", "header_mention": "[['Our Results'], None, ['uni-skip', 'bi-skip', 'combine-skip', 'Med r', 'hierarchical model+emb.', 'composite model+emb.', 'combine+emb.'], ['combine+emb.', 'combine-skip', 'R@1']]", "valid": 1, "image_id": "D17-1254-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results for image-sentence ranking experiments on the COCO dataset. R@K denotes Recall@K (higher is better) and Med r is the median rank (lower is better). (\u2020) taken from Kiros et al. (2015). (\u2217) taken from Karpathy and Fei-Fei (2015). (\u2021) taken from Mao et al. (2015).

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n R@1Med rR@2Med r
Method || uni-skip30.6322.74
Method || bi-skip32.7324.24
Method || combine-skip33.8325.94
Our Results || hierarchical model+emb.32.7325.34
Our Results || composite model+emb.33.8325.74
Our Results || combine+emb.34.4326.64
Task-dependent methods || DVSA*38.4127.43
Task-dependent methods || m-RNN412293
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results for image-sentence ranking experiments on the COCO dataset. R@K denotes Recall@K (higher is better) and Med r is the median rank (lower is better). ($\\dagger$) taken from Kiros et al. (2015). ($\\ast$) taken from Karpathy and Fei-Fei (2015). ($\\ddagger$) taken from Mao et al. (2015).}\n\\begin{tabular}{lrrrr}\n\\toprule\n & R@1 & Med r & R@2 & Med r.1 \\\\\n\\midrule\nMethod $\\parallel$ uni-skip & 30.6 & 3 & 22.7 & 4 \\\\\nMethod $\\parallel$ bi-skip & 32.7 & 3 & 24.2 & 4 \\\\\nMethod $\\parallel$ combine-skip & 33.8 & 3 & 25.9 & 4 \\\\\nOur Results $\\parallel$ hierarchical model+emb. & 32.7 & 3 & 25.3 & 4 \\\\\nOur Results $\\parallel$ composite model+emb. & 33.8 & 3 & 25.7 & 4 \\\\\nOur Results $\\parallel$ combine+emb. & 34.4 & 3 & 26.6 & 4 \\\\\nTask-dependent methods $\\parallel$ DVSA* & 38.4 & 1 & 27.4 & 3 \\\\\nTask-dependent methods $\\parallel$ m-RNN & 41 & 2 & 29 & 3 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Results for image-sentence ranking experiments on the COCO dataset. R@K denotes Recall@K (higher is better) and Med r is the median rank (lower is better). (\u2020) taken from Kiros et al. (2015). (\u2217) taken from Karpathy and Fei-Fei (2015). (\u2021) taken from Mao et al. (2015).\n
\n \n R@1\n \n Med r\n \n R@2\n \n Med r\n
\n Method || uni-skip\n \n 30.6\n \n 3\n \n 22.7\n \n 4\n
\n Method || bi-skip\n \n 32.7\n \n 3\n \n 24.2\n \n 4\n
\n Method || combine-skip\n \n 33.8\n \n 3\n \n 25.9\n \n 4\n
\n Our Results || hierarchical model+emb.\n \n 32.7\n \n 3\n \n 25.3\n \n 4\n
\n Our Results || composite model+emb.\n \n 33.8\n \n 3\n \n 25.7\n \n 4\n
\n Our Results || combine+emb.\n \n 34.4\n \n 3\n \n 26.6\n \n 4\n
\n Task-dependent methods || DVSA*\n \n 38.4\n \n 1\n \n 27.4\n \n 3\n
\n Task-dependent methods || m-RNN\n \n 41\n \n 2\n \n 29\n \n 3\n
\n" }, { "instance_id": "a041eaeb-0a31-4c5a-aeeb-d7cbdfb1ed81", "table_id_paper": "D17-1267table_4", "caption": "Cognate clustering results on the Algonquian dataset (in %) with subsets of features.", "row_header_level": 2, "row_headers": [ [ "Features", "Phonetic only" ], [ "Features", "+ Definitions" ], [ "Features", "+ WordNet" ], [ "Features", "+ Word Vectors" ] ], "column_header_level": 1, "column_headers": [ [ "Found Sets" ], [ "Purity" ] ], "contents": [ [ "52.0 (36.3)", "70.2" ], [ "57.4 (41.7)", "68.4" ], [ "61.9 (46.9)", "68.1" ], [ "66.2 (51.3)", "66.5" ] ], "metrics_loc": "column", "metrics_type": "['Found Sets', 'Purity']", "target_entity": "['Phonetic only']", "table_html": "
Found Sets Purity
Features || Phonetic only 52.0 (36.3) 70.2
Features || + Definitions 57.4 (41.7) 68.4
Features || + WordNet 61.9 (46.9) 68.1
Features || + Word Vectors 66.2 (51.3) 66.5
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D17-1267", "page_no": 8, "dir": "emnlp2017", "description": "Table 4 shows the results. The phonetic features alone are sufficient to detect just over half of the cognate sets. Each successive variant substantially improves the recall at a cost of slightly lower precision. The full feature set yields a 27% relative increase in the number of found sets over the phonetic-only variant, with only a 5% drop in cluster purity.", "class_sentence": "[1, 1, 1, 1]", "sentences": "['Table 4 shows the results.', 'The phonetic features alone are sufficient to detect just over half of the cognate sets.', 'Each successive variant substantially improves the recall at a cost of slightly lower precision.', 'The full feature set yields a 27% relative increase in the number of found sets over the phonetic-only variant, with only a 5% drop in cluster purity.']", "header_mention": "[None, ['Phonetic only', 'Purity'], ['+ Definitions', '+ WordNet', '+ Word Vectors', 'Purity'], ['Phonetic only', 'Found Sets', 'Purity']]", "valid": 1, "image_id": "D17-1267-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Cognate clustering results on the Algonquian dataset (in %) with subsets of features.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Found SetsPurity
Features || Phonetic only52.0 (36.3)70.2
Features || + Definitions57.4 (41.7)68.4
Features || + WordNet61.9 (46.9)68.1
Features || + Word Vectors66.2 (51.3)66.5
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Cognate clustering results on the Algonquian dataset (in \\%) with subsets of features.}\n\\begin{tabular}{llr}\n\\toprule\n & Found Sets & Purity \\\\\n\\midrule\nFeatures $\\parallel$ Phonetic only & 52.0 (36.3) & 70.2 \\\\\nFeatures $\\parallel$ + Definitions & 57.4 (41.7) & 68.4 \\\\\nFeatures $\\parallel$ + WordNet & 61.9 (46.9) & 68.1 \\\\\nFeatures $\\parallel$ + Word Vectors & 66.2 (51.3) & 66.5 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Cognate clustering results on the Algonquian dataset (in %) with subsets of features.\n
\n \n Found Sets\n \n Purity\n
\n Features || Phonetic only\n \n 52.0 (36.3)\n \n 70.2\n
\n Features || + Definitions\n \n 57.4 (41.7)\n \n 68.4\n
\n Features || + WordNet\n \n 61.9 (46.9)\n \n 68.1\n
\n Features || + Word Vectors\n \n 66.2 (51.3)\n \n 66.5\n
\n" }, { "instance_id": "455406ba-3ee5-4f13-afe2-e9a6d1721e45", "table_id_paper": "D17-1304table_1", "caption": "Results on NIST Chinese-to-English Translation Task. \u201c*\u201d indicates statistically significant better than \u201cSennrich-deponly\u201d at p-value < 0.05 and \u201c**\u201d at p-value < 0.01. AVG = average BLEU scores for test sets.", "row_header_level": 2, "row_headers": [ [ "System", "PBSMT" ], [ "System", "AttNMT" ], [ "System", "Sennrich-deponly" ], [ "System", "SDRNMT-1" ], [ "System", "SDRNMT-2" ] ], "column_header_level": 1, "column_headers": [ [ "Dev (MT02)" ], [ "MT03" ], [ "MT04" ], [ "MT05" ], [ "MT06" ], [ "MT08" ], [ "AVG" ] ], "contents": [ [ "33.15", "31.02", "33.78", "30.33", "29.62", "23.53", "29.66" ], [ "36.31", "34.02", "37.11", "32.86", "32.54", "25.44", "32.4" ], [ "36.68", "34.51", "38.09", "33.37", "32.96", "26.96", "32.98" ], [ "36.88", "34.98*", "38.14", "34.61", "33.58", "27.06", "33.32" ], [ "37.34", "35.91**", "38.73*", "34.18**", "33.76**", "27.64*", "34.04" ] ], "metrics_loc": "column", "metrics_type": "['BLEU', 'BLEU', 'BLEU', 'BLEU', 'BLEU', 'BLEU', 'BLEU']", "target_entity": "['SDRNMT-2']", "table_html": "
Dev (MT02) MT03 MT04 MT05 MT06 MT08 AVG
System || PBSMT 33.15 31.02 33.78 30.33 29.62 23.53 29.66
System || AttNMT 36.31 34.02 37.11 32.86 32.54 25.44 32.4
System || Sennrich-deponly 36.68 34.51 38.09 33.37 32.96 26.96 32.98
System || SDRNMT-1 36.88 34.98* 38.14 34.61 33.58 27.06 33.32
System || SDRNMT-2 37.34 35.91** 38.73* 34.18** 33.76** 27.64* 34.04
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "D17-1304", "page_no": 5, "dir": "emnlp2017", "description": "Table 1 shows the translation performances on test sets measured in BLEU score. The AttNMT significantly outperforms PBSMT by 2.74 BLEU points on average, indicating that it is a strong baseline NMT system. The baseline Sennrichdeponly improves the performance over the AttNMT by 0.58 BLEU points on average. This indicates that the proposed source dependency constraint is beneficial for improving the performance of NMT. Moreover, SDRNMT-1 gains improvements of 0.92 and 0.34 BLEU points on average than the AttNMT and Sennrich-deponly. These show that the proposed SDR can more effectively capture source dependency information than vector concatenation. Especially, the proposed SDRNMT2 outperforms the AttNMT and Sennrich-deponly on average by 1.64 and 1.03 BLEU points. These verify that the proposed double-context method is effective for word prediction.", "class_sentence": "[1, 1, 1, 2, 1, 2, 1, 2]", "sentences": "['Table 1 shows the translation performances on test sets measured in BLEU score.', 'The AttNMT significantly outperforms PBSMT by 2.74 BLEU points on average, indicating that it is a strong baseline NMT system.', 'The baseline Sennrichdeponly improves the performance over the AttNMT by 0.58 BLEU points on average.', 'This indicates that the proposed source dependency constraint is beneficial for improving the performance of NMT.', 'Moreover, SDRNMT-1 gains improvements of 0.92 and 0.34 BLEU points on average than the AttNMT and Sennrich-deponly.', 'These show that the proposed SDR can more effectively capture source dependency information than vector concatenation.', 'Especially, the proposed SDRNMT2 outperforms the AttNMT and Sennrich-deponly on average by 1.64 and 1.03 BLEU points.', 'These verify that the proposed double-context method is effective for word prediction.']", "header_mention": "[None, ['AttNMT', 'PBSMT', 'AVG'], ['Sennrich-deponly', 'AttNMT', 'AVG'], ['Sennrich-deponly', 'AttNMT'], ['SDRNMT-1', 'AttNMT', 'Sennrich-deponly', 'AVG'], ['SDRNMT-1', 'AttNMT', 'Sennrich-deponly'], ['SDRNMT-2', 'AttNMT', 'Sennrich-deponly', 'AVG'], ['SDRNMT-2', 'AttNMT', 'Sennrich-deponly']]", "valid": 1, "image_id": "D17-1304-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results on NIST Chinese-to-English Translation Task. \u201c*\u201d indicates statistically significant better than \u201cSennrich-deponly\u201d at p-value < 0.05 and \u201c**\u201d at p-value < 0.01. AVG = average BLEU scores for test sets.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Dev (MT02)MT03MT04MT05MT06MT08AVG
System || PBSMT33.1531.0233.7830.3329.6223.5329.66
System || AttNMT36.3134.0237.1132.8632.5425.4432.4
System || Sennrich-deponly36.6834.5138.0933.3732.9626.9632.98
System || SDRNMT-136.8834.98*38.1434.6133.5827.0633.32
System || SDRNMT-237.3435.91**38.73*34.18**33.76**27.64*34.04
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results on NIST Chinese-to-English Translation Task. \u201c*\u201d indicates statistically significant better than \u201cSennrich-deponly\u201d at p-value $<$ 0.05 and \u201c**\u201d at p-value $<$ 0.01. AVG $=$ average BLEU scores for test sets.}\n\\begin{tabular}{lrlllllr}\n\\toprule\n & Dev (MT02) & MT03 & MT04 & MT05 & MT06 & MT08 & AVG \\\\\n\\midrule\nSystem $\\parallel$ PBSMT & 33.15 & 31.02 & 33.78 & 30.33 & 29.62 & 23.53 & 29.66 \\\\\nSystem $\\parallel$ AttNMT & 36.31 & 34.02 & 37.11 & 32.86 & 32.54 & 25.44 & 32.4 \\\\\nSystem $\\parallel$ Sennrich-deponly & 36.68 & 34.51 & 38.09 & 33.37 & 32.96 & 26.96 & 32.98 \\\\\nSystem $\\parallel$ SDRNMT-1 & 36.88 & 34.98* & 38.14 & 34.61 & 33.58 & 27.06 & 33.32 \\\\\nSystem $\\parallel$ SDRNMT-2 & 37.34 & 35.91** & 38.73* & 34.18** & 33.76** & 27.64* & 34.04 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Results on NIST Chinese-to-English Translation Task. \u201c*\u201d indicates statistically significant better than \u201cSennrich-deponly\u201d at p-value < 0.05 and \u201c**\u201d at p-value < 0.01. AVG = average BLEU scores for test sets.\n
\n \n Dev (MT02)\n \n MT03\n \n MT04\n \n MT05\n \n MT06\n \n MT08\n \n AVG\n
\n System || PBSMT\n \n 33.15\n \n 31.02\n \n 33.78\n \n 30.33\n \n 29.62\n \n 23.53\n \n 29.66\n
\n System || AttNMT\n \n 36.31\n \n 34.02\n \n 37.11\n \n 32.86\n \n 32.54\n \n 25.44\n \n 32.4\n
\n System || Sennrich-deponly\n \n 36.68\n \n 34.51\n \n 38.09\n \n 33.37\n \n 32.96\n \n 26.96\n \n 32.98\n
\n System || SDRNMT-1\n \n 36.88\n \n 34.98*\n \n 38.14\n \n 34.61\n \n 33.58\n \n 27.06\n \n 33.32\n
\n System || SDRNMT-2\n \n 37.34\n \n 35.91**\n \n 38.73*\n \n 34.18**\n \n 33.76**\n \n 27.64*\n \n 34.04\n
\n" }, { "instance_id": "f85fb334-1443-4f23-bf62-69834ad01fd4", "table_id_paper": "D17-1309table_4", "caption": "Segmentation results. Explicit bigrams are useful.", "row_header_level": 2, "row_headers": [ [ "Model", "Zhang et al. (2016)" ], [ "Model", "Zhang et al. (2016)-combo" ], [ "Model", "Small FF 64 dim" ], [ "Model", "Small FF 256 dim" ], [ "Model", "Small FF 64 dim bigrams" ] ], "column_header_level": 1, "column_headers": [ [ "Accuracy" ], [ "Size" ] ], "contents": [ [ "95.01", "\u2212" ], [ "95.95", "\u2212" ], [ "94.24", "846KB" ], [ "94.16", "3.2MB" ], [ "95.18", "2.0MB" ] ], "metrics_loc": "column", "metrics_type": "['Accuracy', 'Size']", "target_entity": "['Small FF 64 dim bigrams']", "table_html": "
Accuracy Size
Model || Zhang et al. (2016) 95.01 \u2212
Model || Zhang et al. (2016)-combo 95.95 \u2212
Model || Small FF 64 dim 94.24 846KB
Model || Small FF 256 dim 94.16 3.2MB
Model || Small FF 64 dim bigrams 95.18 2.0MB
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D17-1309", "page_no": 4, "dir": "emnlp2017", "description": "The results in Table 4 show that spending our memory budget on small bigram embeddings is more effective than on larger character embeddings, in terms of both accuracy and model size. Our model featuring bigrams runs at 110KB of text per second, or 39k tokens/second.", "class_sentence": "[1, 1]", "sentences": "['The results in Table 4 show that spending our memory budget on small bigram embeddings is more effective than on larger character embeddings, in terms of both accuracy and model size.', 'Our model featuring bigrams runs at 110KB of text per second, or 39k tokens/second.']", "header_mention": "[['Small FF 64 dim bigrams', 'Accuracy', 'Size'], ['Small FF 64 dim bigrams']]", "valid": 1, "image_id": "D17-1309-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Segmentation results. Explicit bigrams are useful.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n AccuracySize
Model || Zhang et al. (2016)95.01\u2212
Model || Zhang et al. (2016)-combo95.95\u2212
Model || Small FF 64 dim94.24846KB
Model || Small FF 256 dim94.163.2MB
Model || Small FF 64 dim bigrams95.182.0MB
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Segmentation results. Explicit bigrams are useful.}\n\\begin{tabular}{lrl}\n\\toprule\n & Accuracy & Size \\\\\n\\midrule\nModel $\\parallel$ Zhang et al. (2016) & 95.01 & - \\\\\nModel $\\parallel$ Zhang et al. (2016)-combo & 95.95 & - \\\\\nModel $\\parallel$ Small FF 64 dim & 94.24 & 846KB \\\\\nModel $\\parallel$ Small FF 256 dim & 94.16 & 3.2MB \\\\\nModel $\\parallel$ Small FF 64 dim bigrams & 95.18 & 2.0MB \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Segmentation results. Explicit bigrams are useful.\n
\n \n Accuracy\n \n Size\n
\n Model || Zhang et al. (2016)\n \n 95.01\n \n \u2212\n
\n Model || Zhang et al. (2016)-combo\n \n 95.95\n \n \u2212\n
\n Model || Small FF 64 dim\n \n 94.24\n \n 846KB\n
\n Model || Small FF 256 dim\n \n 94.16\n \n 3.2MB\n
\n Model || Small FF 64 dim bigrams\n \n 95.18\n \n 2.0MB\n
\n" }, { "instance_id": "94333a40-201d-4c8e-a1d9-3f8a4ccbcc68", "table_id_paper": "D18-1033table_5", "caption": "Performance on monolingual word similarity computation with seed lexicon size 6000.", "row_header_level": 2, "row_headers": [ [ "Method", "BiLex" ], [ "Method", "CLSP-WR" ], [ "Method", "CLSP-SE" ] ], "column_header_level": 2, "column_headers": [ [ "Chinese (source)", "WS-240" ], [ "Chinese (source)", "WS-297" ], [ "English (target)", "WS-353" ], [ "English (target)", "SL-999" ] ], "contents": [ [ "60.36", "62.17", "60.46", "27.22" ], [ "61.27", "65.25", "60.46", "27.22" ], [ "60.84", "65.62", "62.47", "28.79" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy', 'accuracy']", "target_entity": "['CLSP-WR', 'CLSP-SE']", "table_html": "
Chinese (source) || WS-240 Chinese (source) || WS-297 English (target) || WS-353 English (target) || SL-999
Method || BiLex 60.36 62.17 60.46 27.22
Method || CLSP-WR 61.27 65.25 60.46 27.22
Method || CLSP-SE 60.84 65.62 62.47 28.79
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "D18-1033", "page_no": 9, "dir": "emnlp2018", "description": "Table 5 shows the results of monolingual word similarity computation on four datasets. From the table, we find that:. (1) Our models perform better than BiLex on both Chinese word similarity datasets. It signifies incorporating sememe information helps learn better monolingual embeddings;. (2) CLSP-WR model does not enhance English word similarity results but CLSPSE model does. It is because CLSP-WR model only post-processes Chinese word embeddings and keeps English word embeddings unchanged while CLSP-SE model undertakes bilingual alignment and sememe information incorporation together, which makes English word embeddings improve with Chinese word embeddings.", "class_sentence": "[1, 1, 1, 2, 1, 2]", "sentences": "['Table 5 shows the results of monolingual word similarity computation on four datasets.', 'From the table, we find that:.', '(1) Our models perform better than BiLex on both Chinese word similarity datasets.', 'It signifies incorporating sememe information helps learn better monolingual embeddings;.', '(2) CLSP-WR model does not enhance English word similarity results but CLSPSE model does.', 'It is because CLSP-WR model only post-processes Chinese word embeddings and keeps English word embeddings unchanged while CLSP-SE model undertakes bilingual alignment and sememe information incorporation together, which makes English word embeddings improve with Chinese word embeddings.']", "header_mention": "[None, None, ['BiLex', 'CLSP-WR', 'CLSP-SE', 'Chinese (source)', 'WS-240', 'WS-297'], None, ['CLSP-WR', 'CLSP-SE', 'English (target)', 'WS-353', 'SL-999'], ['CLSP-WR', 'CLSP-SE', 'Chinese (source)', 'English (target)']]", "valid": 1, "image_id": "D18-1033-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance on monolingual word similarity computation with seed lexicon size 6000.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Chinese (source) || WS-240Chinese (source) || WS-297English (target) || WS-353English (target) || SL-999
Method || BiLex60.3662.1760.4627.22
Method || CLSP-WR61.2765.2560.4627.22
Method || CLSP-SE60.8465.6262.4728.79
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance on monolingual word similarity computation with seed lexicon size 6000.}\n\\begin{tabular}{lrrrr}\n\\toprule\n & Chinese (source) $\\parallel$ WS-240 & Chinese (source) $\\parallel$ WS-297 & English (target) $\\parallel$ WS-353 & English (target) $\\parallel$ SL-999 \\\\\n\\midrule\nMethod $\\parallel$ BiLex & 60.36 & 62.17 & 60.46 & 27.22 \\\\\nMethod $\\parallel$ CLSP-WR & 61.27 & 65.25 & 60.46 & 27.22 \\\\\nMethod $\\parallel$ CLSP-SE & 60.84 & 65.62 & 62.47 & 28.79 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Performance on monolingual word similarity computation with seed lexicon size 6000.\n
\n \n Chinese (source) || WS-240\n \n Chinese (source) || WS-297\n \n English (target) || WS-353\n \n English (target) || SL-999\n
\n Method || BiLex\n \n 60.36\n \n 62.17\n \n 60.46\n \n 27.22\n
\n Method || CLSP-WR\n \n 61.27\n \n 65.25\n \n 60.46\n \n 27.22\n
\n Method || CLSP-SE\n \n 60.84\n \n 65.62\n \n 62.47\n \n 28.79\n
\n" }, { "instance_id": "d53e946c-c08c-4dfb-abc5-06d17810eeff", "table_id_paper": "D18-1049table_4", "caption": "Comparison with Transformer on FrenchEnglish translation task. The evaluation metric is caseinsensitive BLEU score.", "row_header_level": 2, "row_headers": [ [ "Method", "Transformer" ], [ "Method", "this work" ] ], "column_header_level": 1, "column_headers": [ [ "Dev" ], [ "Test" ] ], "contents": [ [ "29.42", "35.15" ], [ "30.40", "36.04" ] ], "metrics_loc": "column", "metrics_type": "['BLEU', 'BLEU']", "target_entity": "['this work']", "table_html": "
Dev Test
Method || Transformer 29.42 35.15
Method || this work 30.40 36.04
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D18-1049", "page_no": 7, "dir": "emnlp2018", "description": "Table 4 shows that our model also outperforms Transformer by 0.89 BLEU points on French-English translation task.", "class_sentence": "[1]", "sentences": "['Table 4 shows that our model also outperforms Transformer by 0.89 BLEU points on French-English translation task.']", "header_mention": "[['this work', 'Transformer', 'Test']]", "valid": 1, "image_id": "D18-1049-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison with Transformer on FrenchEnglish translation task. The evaluation metric is caseinsensitive BLEU score.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n DevTest
Method || Transformer29.4235.15
Method || this work30.4036.04
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison with Transformer on FrenchEnglish translation task. The evaluation metric is caseinsensitive BLEU score.}\n\\begin{tabular}{lrr}\n\\toprule\n & Dev & Test \\\\\n\\midrule\nMethod $\\parallel$ Transformer & 29.42 & 35.15 \\\\\nMethod $\\parallel$ this work & 30.4 & 36.04 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Comparison with Transformer on FrenchEnglish translation task. The evaluation metric is caseinsensitive BLEU score.\n
\n \n Dev\n \n Test\n
\n Method || Transformer\n \n 29.42\n \n 35.15\n
\n Method || this work\n \n 30.40\n \n 36.04\n
\n" }, { "instance_id": "18bb0280-54f0-4524-b15f-48ab43399a85", "table_id_paper": "D18-1054table_2", "caption": "Performance of various models on NarrativeQA dataset (SS=sample size \u2261 number of relevant sentences)", "row_header_level": 2, "row_headers": [ [ "Model", "Seq2Seq" ], [ "Model", "ASR" ], [ "Model", "BiDAF" ], [ "Model", "MRU (Tay et al. 2018)" ], [ "Model", "Baseline 1 (SS=5)" ], [ "Model", "Baseline 2 (SS=5)" ], [ "Model", "ConZNet (SS=1)" ], [ "Model", "ConZNet (SS=3)" ], [ "Model", "ConZNet (SS=5)" ], [ "Model", "ConZNet (SS=7)" ] ], "column_header_level": 1, "column_headers": [ [ "BLEU-1" ], [ "BLEU-4" ], [ "ROUGE-L" ], [ "METEOR" ] ], "contents": [ [ "15.89", "1.26", "13.15", "4.08" ], [ "23.20", "6.39", "22.26", "7.77" ], [ "33.72", "15.53", "36.30", "15.38" ], [ "36.55", "19.79", "41.44", "17.87" ], [ "30.22", "14.43", "34.40", "13.36" ], [ "39.35", "20.17", "43.36", "18.01" ], [ "28.97", "16.70", "36.02", "12.39" ], [ "36.21", "19.33", "41.23", "17.77" ], [ "42.76", "22.49", "46.67", "19.24" ], [ "40.80", "21.14", "44.01", "18.67" ] ], "metrics_loc": "column", "metrics_type": "['BLEU-1', 'BLEU-4', 'ROUGE-L', 'METEOR']", "target_entity": "['ConZNet (SS=1)', 'ConZNet (SS=3)', 'ConZNet (SS=5)', 'ConZNet (SS=7)']", "table_html": "
BLEU-1 BLEU-4 ROUGE-L METEOR
Model || Seq2Seq 15.89 1.26 13.15 4.08
Model || ASR 23.20 6.39 22.26 7.77
Model || BiDAF 33.72 15.53 36.30 15.38
Model || MRU (Tay et al. 2018) 36.55 19.79 41.44 17.87
Model || Baseline 1 (SS=5) 30.22 14.43 34.40 13.36
Model || Baseline 2 (SS=5) 39.35 20.17 43.36 18.01
Model || ConZNet (SS=1) 28.97 16.70 36.02 12.39
Model || ConZNet (SS=3) 36.21 19.33 41.23 17.77
Model || ConZNet (SS=5) 42.76 22.49 46.67 19.24
Model || ConZNet (SS=7) 40.80 21.14 44.01 18.67
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1054", "page_no": 5, "dir": "emnlp2018", "description": "3.4 Results. Table 2 shows the performance of various models on NarrativeQA. It can be noted that our model with sample size 5 (choosing 5 relevant sentences) outperforms the best ROUGE-L score available so far by 12.62% compared to Tay et al. (2018). The low performance of Baseline 1 shows that the hybrid approach (ConZNet) for generating words from a fixed vocabulary as well as copying words from the document is better suited than span prediction models (Seq2Seq, ASR, BiDAF, MRU). To validate the importance of finding relevant sentences in contrast to using an entire document for answer generation, we experimented with sample sizes beyond 5. The performance of our model gradually dropped from sample size 7 onwards. This result shows evidence that only a few relevant sentences are sufficient to answer a question. We also experimented with various sample sizes to see the effect of intra sentence relations for answer generation. The performance of the model improved dramatically with sample sizes 3 and 5 compared to the sample size of 1. These results show that the importance of selecting multiple relevant sentences for generating an answer. In addition, the low performance of Baseline 2 indicates that just selecting multiple sentences is not enough, they should also be related to each other. This result points out that the self-attention mechanism in the Context zoom layer is an important component to identify related relevant sentences.", "class_sentence": "[2, 1, 1, 1, 2, 1, 2, 2, 1, 2, 1, 2]", "sentences": "['3.4 Results.', 'Table 2 shows the performance of various models on NarrativeQA.', 'It can be noted that our model with sample size 5 (choosing 5 relevant sentences) outperforms the best ROUGE-L score available so far by 12.62% compared to Tay et al. (2018).', 'The low performance of Baseline 1 shows that the hybrid approach (ConZNet) for generating words from a fixed vocabulary as well as copying words from the document is better suited than span prediction models (Seq2Seq, ASR, BiDAF, MRU).', 'To validate the importance of finding relevant sentences in contrast to using an entire document for answer generation, we experimented with sample sizes beyond 5.', 'The performance of our model gradually dropped from sample size 7 onwards.', 'This result shows evidence that only a few relevant sentences are sufficient to answer a question.', 'We also experimented with various sample sizes to see the effect of intra sentence relations for answer generation.', 'The performance of the model improved dramatically with sample sizes 3 and 5 compared to the sample size of 1.', 'These results show that the importance of selecting multiple relevant sentences for generating an answer.', 'In addition, the low performance of Baseline 2 indicates that just selecting multiple sentences is not enough, they should also be related to each other.', 'This result points out that the self-attention mechanism in the Context zoom layer is an important component to identify related relevant sentences.']", "header_mention": "[None, ['Seq2Seq', 'ASR', 'BiDAF', 'MRU (Tay et al. 2018)', 'Baseline 1 (SS=5)', 'Baseline 2 (SS=5)', 'ConZNet (SS=1)', 'ConZNet (SS=3)', 'ConZNet (SS=5)', 'ConZNet (SS=7)'], ['ConZNet (SS=5)', 'MRU (Tay et al. 2018)', 'ROUGE-L'], ['Baseline 1 (SS=5)', 'Seq2Seq', 'ASR', 'BiDAF', 'MRU (Tay et al. 2018)'], ['ConZNet (SS=7)'], ['ConZNet (SS=7)'], None, ['ConZNet (SS=1)', 'ConZNet (SS=3)', 'ConZNet (SS=5)', 'ConZNet (SS=7)'], ['ConZNet (SS=1)', 'ConZNet (SS=3)', 'ConZNet (SS=5)'], None, ['Baseline 2 (SS=5)'], None]", "valid": 1, "image_id": "D18-1054-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance of various models on NarrativeQA dataset (SS=sample size \u2261 number of relevant sentences)

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n BLEU-1BLEU-4ROUGE-LMETEOR
Model || Seq2Seq15.891.2613.154.08
Model || ASR23.206.3922.267.77
Model || BiDAF33.7215.5336.3015.38
Model || MRU (Tay et al. 2018)36.5519.7941.4417.87
Model || Baseline 1 (SS=5)30.2214.4334.4013.36
Model || Baseline 2 (SS=5)39.3520.1743.3618.01
Model || ConZNet (SS=1)28.9716.7036.0212.39
Model || ConZNet (SS=3)36.2119.3341.2317.77
Model || ConZNet (SS=5)42.7622.4946.6719.24
Model || ConZNet (SS=7)40.8021.1444.0118.67
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance of various models on NarrativeQA dataset (SS$=$sample size $\\equiv$ number of relevant sentences)}\n\\begin{tabular}{lrrrr}\n\\toprule\n & BLEU-1 & BLEU-4 & ROUGE-L & METEOR \\\\\n\\midrule\nModel $\\parallel$ Seq2Seq & 15.89 & 1.26 & 13.15 & 4.08 \\\\\nModel $\\parallel$ ASR & 23.2 & 6.39 & 22.26 & 7.77 \\\\\nModel $\\parallel$ BiDAF & 33.72 & 15.53 & 36.3 & 15.38 \\\\\nModel $\\parallel$ MRU (Tay et al. 2018) & 36.55 & 19.79 & 41.44 & 17.87 \\\\\nModel $\\parallel$ Baseline 1 (SS$=$5) & 30.22 & 14.43 & 34.4 & 13.36 \\\\\nModel $\\parallel$ Baseline 2 (SS$=$5) & 39.35 & 20.17 & 43.36 & 18.01 \\\\\nModel $\\parallel$ ConZNet (SS$=$1) & 28.97 & 16.7 & 36.02 & 12.39 \\\\\nModel $\\parallel$ ConZNet (SS$=$3) & 36.21 & 19.33 & 41.23 & 17.77 \\\\\nModel $\\parallel$ ConZNet (SS$=$5) & 42.76 & 22.49 & 46.67 & 19.24 \\\\\nModel $\\parallel$ ConZNet (SS$=$7) & 40.8 & 21.14 & 44.01 & 18.67 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Performance of various models on NarrativeQA dataset (SS=sample size \u2261 number of relevant sentences)\n
\n \n BLEU-1\n \n BLEU-4\n \n ROUGE-L\n \n METEOR\n
\n Model || Seq2Seq\n \n 15.89\n \n 1.26\n \n 13.15\n \n 4.08\n
\n Model || ASR\n \n 23.20\n \n 6.39\n \n 22.26\n \n 7.77\n
\n Model || BiDAF\n \n 33.72\n \n 15.53\n \n 36.30\n \n 15.38\n
\n Model || MRU (Tay et al. 2018)\n \n 36.55\n \n 19.79\n \n 41.44\n \n 17.87\n
\n Model || Baseline 1 (SS=5)\n \n 30.22\n \n 14.43\n \n 34.40\n \n 13.36\n
\n Model || Baseline 2 (SS=5)\n \n 39.35\n \n 20.17\n \n 43.36\n \n 18.01\n
\n Model || ConZNet (SS=1)\n \n 28.97\n \n 16.70\n \n 36.02\n \n 12.39\n
\n Model || ConZNet (SS=3)\n \n 36.21\n \n 19.33\n \n 41.23\n \n 17.77\n
\n Model || ConZNet (SS=5)\n \n 42.76\n \n 22.49\n \n 46.67\n \n 19.24\n
\n Model || ConZNet (SS=7)\n \n 40.80\n \n 21.14\n \n 44.01\n \n 18.67\n
\n" }, { "instance_id": "fad7b4e8-a0cc-4f97-9e53-01a85da9e783", "table_id_paper": "D18-1058table_3", "caption": "Published results of other models on the SemEval2012 Task 2 dataset.", "row_header_level": 2, "row_headers": [ [ "Model", "Rink and Harabagiu (2012)" ], [ "Model", "Mikolov et al. (2013b)" ], [ "Model", "Levy and Goldberg (2014)" ], [ "Model", "Zhila et al. (2013)" ], [ "Model", "Iacobacci et al. (2015)" ], [ "Model", "Turney (2013)" ], [ "Model", "VecOff" ], [ "Model", "LRA" ], [ "Model", "NLRA" ], [ "Model", "NLRA+VecOff" ] ], "column_header_level": 1, "column_headers": [ [ " Accuracy" ], [ "Correlation" ] ], "contents": [ [ "0.394", "0.229" ], [ "0.418", "0.275" ], [ "0.452", "-" ], [ "0.452", "0.353" ], [ "0.459", "0.358" ], [ "0.472", "0.408" ], [ "0.443", "0.321" ], [ "0.415", "0.264" ], [ "0.453", "0.36" ], [ "0.475", "0.391" ] ], "metrics_loc": "column", "metrics_type": "['Accuracy', 'Correlation']", "target_entity": "['NLRA', 'NLRA+VecOff']", "table_html": "
Accuracy Correlation
Model || Rink and Harabagiu (2012) 0.394 0.229
Model || Mikolov et al. (2013b) 0.418 0.275
Model || Levy and Goldberg (2014) 0.452 -
Model || Zhila et al. (2013) 0.452 0.353
Model || Iacobacci et al. (2015) 0.459 0.358
Model || Turney (2013) 0.472 0.408
Model || VecOff 0.443 0.321
Model || LRA 0.415 0.264
Model || NLRA 0.453 0.36
Model || NLRA+VecOff 0.475 0.391
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D18-1058", "page_no": 4, "dir": "emnlp2018", "description": "4.5 Comparison to other systems. We compared the results of our models to other published results. Table 3 displays those results. Rink and Harabagiu (2012) is the pattern-based model with naive Bayes. Mikolov et al. (2013b), Levy and Goldberg (2014), and Iacobacci et al. (2015) are the vector offset models. Zhila et al. (2013) is the model composed of various features. Turney (2013) extracts the statistical features of two word pairs from a word-context co-occurrence matrix and trains the classifier with additional semantic relational data to assign a relational similarity for two word pairs. NLRA+VecOff achieved a competitive performance to the state-of-the-art method of Turney (2013). Note that our method learns unsupervisedly and does not exploit additional resources, and the method of Turney (2013) cannot obtain the distributed representation of word pairs. A work similar to ours, Bollegala et al. (2015), represented lexico-syntactic patterns as the vector offset of co-occurring word pairs and updated the vector offsets of word pairs such that word pairs that co-occur in similar patterns have similar offsets. They evaluated their model on all 79 semantic relations of the dataset and achieved 0.449 accuracy. In their setting, NLRA+VecOff achieved 0.47 accuracy, outperforming their model.", "class_sentence": "['2', '2', '1', '2', '2', '2', '2', '1', '2', '0', '2', '1']", "sentences": "['4.5 Comparison to other systems.', 'We compared the results of our models to other published results.', 'Table 3 displays those results.', 'Rink and Harabagiu (2012) is the pattern-based model with naive Bayes.', 'Mikolov et al. (2013b), Levy and Goldberg (2014), and Iacobacci et al. (2015) are the vector offset models.', 'Zhila et al. (2013) is the model composed of various features.', 'Turney (2013) extracts the statistical features of two word pairs from a word-context co-occurrence matrix and trains the classifier with additional semantic relational data to assign a relational similarity for two word pairs.', 'NLRA+VecOff achieved a competitive performance to the state-of-the-art method of Turney (2013).', 'Note that our method learns unsupervisedly and does not exploit additional resources, and the method of Turney (2013) cannot obtain the distributed representation of word pairs.', 'A work similar to ours, Bollegala et al. (2015), represented lexico-syntactic patterns as the vector offset of co-occurring word pairs and updated the vector offsets of word pairs such that word pairs that co-occur in similar patterns have similar offsets.', 'They evaluated their model on all 79 semantic relations of the dataset and achieved 0.449 accuracy.', 'In their setting, NLRA+VecOff achieved 0.47 accuracy, outperforming their model.']", "header_mention": "[None, ['Rink and Harabagiu (2012)', 'Mikolov et al. (2013b)', 'Levy and Goldberg (2014)', 'Zhila et al. (2013)', 'Iacobacci et al. (2015)', 'Turney (2013)', 'VecOff', 'LRA', 'NLRA'], None, ['Rink and Harabagiu (2012)'], ['Mikolov et al. (2013b)', 'Levy and Goldberg (2014)', 'Iacobacci et al. (2015)'], ['Zhila et al. (2013)'], ['Turney (2013)'], ['NLRA+VecOff', 'Turney (2013)'], ['NLRA+VecOff', 'Turney (2013)'], None, None, ['NLRA+VecOff']]", "valid": 1, "image_id": "D18-1058-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Published results of other models on the SemEval2012 Task 2 dataset.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n AccuracyCorrelation
Model || Rink and Harabagiu (2012)0.3940.229
Model || Mikolov et al. (2013b)0.4180.275
Model || Levy and Goldberg (2014)0.452-
Model || Zhila et al. (2013)0.4520.353
Model || Iacobacci et al. (2015)0.4590.358
Model || Turney (2013)0.4720.408
Model || VecOff0.4430.321
Model || LRA0.4150.264
Model || NLRA0.4530.36
Model || NLRA+VecOff0.4750.391
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Published results of other models on the SemEval2012 Task 2 dataset.}\n\\begin{tabular}{lrl}\n\\toprule\n & Accuracy & Correlation \\\\\n\\midrule\nModel $\\parallel$ Rink and Harabagiu (2012) & 0.394 & 0.229 \\\\\nModel $\\parallel$ Mikolov et al. (2013b) & 0.418 & 0.275 \\\\\nModel $\\parallel$ Levy and Goldberg (2014) & 0.452 & - \\\\\nModel $\\parallel$ Zhila et al. (2013) & 0.452 & 0.353 \\\\\nModel $\\parallel$ Iacobacci et al. (2015) & 0.459 & 0.358 \\\\\nModel $\\parallel$ Turney (2013) & 0.472 & 0.408 \\\\\nModel $\\parallel$ VecOff & 0.443 & 0.321 \\\\\nModel $\\parallel$ LRA & 0.415 & 0.264 \\\\\nModel $\\parallel$ NLRA & 0.453 & 0.36 \\\\\nModel $\\parallel$ NLRA+VecOff & 0.475 & 0.391 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Published results of other models on the SemEval2012 Task 2 dataset.\n
\n \n Accuracy\n \n Correlation\n
\n Model || Rink and Harabagiu (2012)\n \n 0.394\n \n 0.229\n
\n Model || Mikolov et al. (2013b)\n \n 0.418\n \n 0.275\n
\n Model || Levy and Goldberg (2014)\n \n 0.452\n \n -\n
\n Model || Zhila et al. (2013)\n \n 0.452\n \n 0.353\n
\n Model || Iacobacci et al. (2015)\n \n 0.459\n \n 0.358\n
\n Model || Turney (2013)\n \n 0.472\n \n 0.408\n
\n Model || VecOff\n \n 0.443\n \n 0.321\n
\n Model || LRA\n \n 0.415\n \n 0.264\n
\n Model || NLRA\n \n 0.453\n \n 0.36\n
\n Model || NLRA+VecOff\n \n 0.475\n \n 0.391\n
\n" }, { "instance_id": "b923fff7-4663-4f24-91a6-f4fb1091abd6", "table_id_paper": "D18-1082table_1", "caption": "Automatic evaluation", "row_header_level": 2, "row_headers": [ [ "Model", "S2SA" ], [ "Model", "TAS2S" ], [ "Model", "SPMN500" ], [ "Model", "SPMN1000" ], [ "Model", "DPMN100" ] ], "column_header_level": 1, "column_headers": [ [ "PPL-V" ], [ "PPL-T" ], [ "distinct-1" ], [ "distinct-2" ] ], "contents": [ [ "8.41", "9.05", "0.0809", "0.2110" ], [ "7.38", "7.84", "0.04759", "0.1087" ], [ "7.04", "7.93", "0.06430", "0.1734" ], [ "6.28", "7.72", "0.07347", "0.1909" ], [ "6.45", "7.69", "0.04350", "0.1048" ] ], "metrics_loc": "column", "metrics_type": "['PPL-V', 'PPL-T', 'distinct-1', 'distinct-2']", "target_entity": "['SPMN500', 'SPMN1000', 'DPMN100']", "table_html": "
PPL-V PPL-T distinct-1 distinct-2
Model || S2SA 8.41 9.05 0.0809 0.2110
Model || TAS2S 7.38 7.84 0.04759 0.1087
Model || SPMN500 7.04 7.93 0.06430 0.1734
Model || SPMN1000 6.28 7.72 0.07347 0.1909
Model || DPMN100 6.45 7.69 0.04350 0.1048
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "D18-1082", "page_no": 4, "dir": "emnlp2018", "description": "Automatic Evaluation:. Table 1 shows the automatic evaluation results. We see that both SPMN and DPMN obtain huge improvements over the two baselines in terms of PPL-V and PPL-T. Also, we observe that SPMN1000 outperforms SPMN500 in all the four automatic metrics. Note that each post has the same prototypes provided by SPMN. This is reasonable that the relevant response is more likely to be covered by storing more prototypes in SPMN. As for the DPMN, we can see that DPMN achieves the best performance with only 100 prototypes in terms of PPL-T, compared with the other 4 methods. This suggests that using a retrieval mechanism to incorporate the relevant responses brings more useful information for better response generation. Note that S2SA outperforms the others in terms of distinct-1 and distinct-2. Further human evaluation indicates that many responses generated by S2SA are irrelevant and meaningless, which could inevitably increase the distinct scores.", "class_sentence": "[2, 1, 1, 1, 2, 2, 1, 2, 1, 2]", "sentences": "['Automatic Evaluation:.', 'Table 1 shows the automatic evaluation results.', 'We see that both SPMN and DPMN obtain huge improvements over the two baselines in terms of PPL-V and PPL-T.', 'Also, we observe that SPMN1000 outperforms SPMN500 in all the four automatic metrics.', 'Note that each post has the same prototypes provided by SPMN.', 'This is reasonable that the relevant response is more likely to be covered by storing more prototypes in SPMN.', 'As for the DPMN, we can see that DPMN achieves the best performance with only 100 prototypes in terms of PPL-T, compared with the other 4 methods.', 'This suggests that using a retrieval mechanism to incorporate the relevant responses brings more useful information for better response generation.', 'Note that S2SA outperforms the others in terms of distinct-1 and distinct-2.', 'Further human evaluation indicates that many responses generated by S2SA are irrelevant and meaningless, which could inevitably increase the distinct scores.']", "header_mention": "[None, None, ['SPMN500', 'SPMN1000', 'DPMN100', 'S2SA', 'TAS2S', 'PPL-V', 'PPL-T'], ['SPMN1000', 'SPMN500', 'PPL-V', 'PPL-T', 'distinct-1', 'distinct-2'], ['SPMN500', 'SPMN1000'], ['SPMN500', 'SPMN1000'], ['DPMN100', 'S2SA', 'TAS2S', 'SPMN500', 'SPMN1000', 'PPL-T'], ['DPMN100'], ['S2SA', 'TAS2S', 'SPMN500', 'SPMN1000', 'DPMN100', 'distinct-1', 'distinct-2'], ['S2SA']]", "valid": 1, "image_id": "D18-1082-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Automatic evaluation

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n PPL-VPPL-Tdistinct-1distinct-2
Model || S2SA8.419.050.08090.2110
Model || TAS2S7.387.840.047590.1087
Model || SPMN5007.047.930.064300.1734
Model || SPMN10006.287.720.073470.1909
Model || DPMN1006.457.690.043500.1048
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Automatic evaluation}\n\\begin{tabular}{lrrrr}\n\\toprule\n & PPL-V & PPL-T & distinct-1 & distinct-2 \\\\\n\\midrule\nModel $\\parallel$ S2SA & 8.41 & 9.05 & 0.0809 & 0.211 \\\\\nModel $\\parallel$ TAS2S & 7.38 & 7.84 & 0.04759 & 0.1087 \\\\\nModel $\\parallel$ SPMN500 & 7.04 & 7.93 & 0.0643 & 0.1734 \\\\\nModel $\\parallel$ SPMN1000 & 6.28 & 7.72 & 0.07347 & 0.1909 \\\\\nModel $\\parallel$ DPMN100 & 6.45 & 7.69 & 0.0435 & 0.1048 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Automatic evaluation\n
\n \n PPL-V\n \n PPL-T\n \n distinct-1\n \n distinct-2\n
\n Model || S2SA\n \n 8.41\n \n 9.05\n \n 0.0809\n \n 0.2110\n
\n Model || TAS2S\n \n 7.38\n \n 7.84\n \n 0.04759\n \n 0.1087\n
\n Model || SPMN500\n \n 7.04\n \n 7.93\n \n 0.06430\n \n 0.1734\n
\n Model || SPMN1000\n \n 6.28\n \n 7.72\n \n 0.07347\n \n 0.1909\n
\n Model || DPMN100\n \n 6.45\n \n 7.69\n \n 0.04350\n \n 0.1048\n
\n" }, { "instance_id": "8f66482b-a8ae-4403-8fe7-db4881d54587", "table_id_paper": "D18-1160table_4", "caption": "Unsupervised POS tagging results on WSJ, with fastText vectors as the observed embeddings.", "row_header_level": 2, "row_headers": [ [ "System", "Gaussian HMM" ], [ "System", "Ours (4 layers)" ], [ "System", "Ours (8 layers)" ], [ "System", "Ours (16 layers)" ] ], "column_header_level": 1, "column_headers": [ [ "M-1" ], [ "VM" ] ], "contents": [ [ "72.0", "65.0" ], [ "76.4", "69.3" ], [ "76.8", "69.4" ], [ "67.3", "62.0" ] ], "metrics_loc": "column", "metrics_type": "['M-1', 'VM']", "target_entity": "['Ours (4 layers)', 'Ours (8 layers)', 'Ours (16 layers)']", "table_html": "
M-1 VM
System || Gaussian HMM 72.0 65.0
System || Ours (4 layers) 76.4 69.3
System || Ours (8 layers) 76.8 69.4
System || Ours (16 layers) 67.3 62.0
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D18-1160", "page_no": 8, "dir": "emnlp2018", "description": "These results are summarized in Table 4 and Table 5. While fastText embeddings lead to reduced performance with our model, our approach still yields an improvement over the Gaussian baseline with the new observed embeddings space.", "class_sentence": "[1, 1]", "sentences": "['These results are summarized in Table 4 and Table 5.', 'While fastText embeddings lead to reduced performance with our model, our approach still yields an improvement over the Gaussian baseline with the new observed embeddings space.']", "header_mention": "[None, ['Ours (4 layers)', 'Ours (8 layers)', 'Ours (16 layers)', 'Gaussian HMM']]", "valid": 1, "image_id": "D18-1160-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Unsupervised POS tagging results on WSJ, with fastText vectors as the observed embeddings.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n M-1VM
System || Gaussian HMM72.065.0
System || Ours (4 layers)76.469.3
System || Ours (8 layers)76.869.4
System || Ours (16 layers)67.362.0
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Unsupervised POS tagging results on WSJ, with fastText vectors as the observed embeddings.}\n\\begin{tabular}{lrr}\n\\toprule\n & M-1 & VM \\\\\n\\midrule\nSystem $\\parallel$ Gaussian HMM & 72 & 65 \\\\\nSystem $\\parallel$ Ours (4 layers) & 76.4 & 69.3 \\\\\nSystem $\\parallel$ Ours (8 layers) & 76.8 & 69.4 \\\\\nSystem $\\parallel$ Ours (16 layers) & 67.3 & 62 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Unsupervised POS tagging results on WSJ, with fastText vectors as the observed embeddings.\n
\n \n M-1\n \n VM\n
\n System || Gaussian HMM\n \n 72.0\n \n 65.0\n
\n System || Ours (4 layers)\n \n 76.4\n \n 69.3\n
\n System || Ours (8 layers)\n \n 76.8\n \n 69.4\n
\n System || Ours (16 layers)\n \n 67.3\n \n 62.0\n
\n" }, { "instance_id": "031ba623-e5ef-4bb0-8fe0-0bb4ab6b47af", "table_id_paper": "D18-1166table_2", "caption": "Results for simple and neural models on the test set of RecipeQA dataset.", "row_header_level": 1, "row_headers": [ [ "Hasty Student" ], [ "Impatient Reader (Text only)" ], [ "Impatient Reader (Multimodal)" ] ], "column_header_level": 1, "column_headers": [ [ "Visual Cloze" ], [ "Textual Cloze" ], [ "Visual Coherence" ], [ "Visual Ordering" ] ], "contents": [ [ "27.35", "26.89", "65.80", "40.88" ], [ "-", "28.03", "-", "-" ], [ "27.36", "29.07", "28.08", "26.74" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy', 'accuracy']", "target_entity": "['Hasty Student', 'Impatient Reader (Text only)', 'Impatient Reader (Multimodal)']", "table_html": "
Visual Cloze Textual Cloze Visual Coherence Visual Ordering
Hasty Student 27.35 26.89 65.80 40.88
Impatient Reader (Text only) - 28.03 - -
Impatient Reader (Multimodal) 27.36 29.07 28.08 26.74
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1166", "page_no": 8, "dir": "emnlp2018", "description": "4.3 Baseline Results. We report the performance of the baseline models in Table 2 which indicates the ratio of correct answers against the total questions in the test. For the textual cloze, the comparison between text-only and multimodal Impatient Reader models shows that the additional visual modality helps the model to understand the question better and to provide more accurate answers. While for the cloze style questions, the Impatient Reader outperforms the Hasty student, for the visual coherence and visual ordering style questions Hasty student gives way better results. This demonstrates that better neural models are needed to be able to effectively deal with this kind of questions. Some qualitative examples are provided in the supplementary material.", "class_sentence": "[2, 1, 1, 1, 2, 2]", "sentences": "['4.3 Baseline Results.', 'We report the performance of the baseline models in Table 2 which indicates the ratio of correct answers against the total questions in the test.', 'For the textual cloze, the comparison between text-only and multimodal Impatient Reader models shows that the additional visual modality helps the model to understand the question better and to provide more accurate answers.', 'While for the cloze style questions, the Impatient Reader outperforms the Hasty student, for the visual coherence and visual ordering style questions Hasty student gives way better results.', 'This demonstrates that better neural models are needed to be able to effectively deal with this kind of questions.', 'Some qualitative examples are provided in the supplementary material.']", "header_mention": "[None, ['Hasty Student', 'Impatient Reader (Text only)', 'Impatient Reader (Multimodal)'], ['Textual Cloze', 'Impatient Reader (Text only)', 'Impatient Reader (Multimodal)'], ['Hasty Student', 'Impatient Reader (Text only)', 'Impatient Reader (Multimodal)', 'Visual Cloze', 'Textual Cloze', 'Visual Coherence', 'Visual Ordering'], ['Impatient Reader (Text only)', 'Impatient Reader (Multimodal)'], None]", "valid": 1, "image_id": "D18-1166-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results for simple and neural models on the test set of RecipeQA dataset.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Visual ClozeTextual ClozeVisual CoherenceVisual Ordering
Hasty Student27.3526.8965.8040.88
Impatient Reader (Text only)-28.03--
Impatient Reader (Multimodal)27.3629.0728.0826.74
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results for simple and neural models on the test set of RecipeQA dataset.}\n\\begin{tabular}{llrll}\n\\toprule\n & Visual Cloze & Textual Cloze & Visual Coherence & Visual Ordering \\\\\n\\midrule\nHasty Student & 27.35 & 26.89 & 65.80 & 40.88 \\\\\nImpatient Reader (Text only) & - & 28.03 & - & - \\\\\nImpatient Reader (Multimodal) & 27.36 & 29.07 & 28.08 & 26.74 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Results for simple and neural models on the test set of RecipeQA dataset.\n
\n \n Visual Cloze\n \n Textual Cloze\n \n Visual Coherence\n \n Visual Ordering\n
\n Hasty Student\n \n 27.35\n \n 26.89\n \n 65.80\n \n 40.88\n
\n Impatient Reader (Text only)\n \n -\n \n 28.03\n \n -\n \n -\n
\n Impatient Reader (Multimodal)\n \n 27.36\n \n 29.07\n \n 28.08\n \n 26.74\n
\n" }, { "instance_id": "517a612e-5096-4a4e-9eb1-ec7c0b6c57d9", "table_id_paper": "D18-1176table_1", "caption": "Accuracy scores on the Stanford Natural Language Inference (SNLI) and MultiNLI Mismatched (MNLI) tasks. DME=Dynamic Meta-Embeddings; CDME=Contextualized Dynamic Meta-Embeddings; *=multiple different embedding sets (see Section 4). Number of parameters included in parenthesis. Results averaged over five runs with different random seeds, using a BiLSTM-Max sentence encoder.", "row_header_level": 2, "row_headers": [ [ "Model", "InferSent (Conneau et al., 2017)" ], [ "Model", "NSE (Munkhdalai and Yu, 2017)" ], [ "Model", "G-TreeLSTM (Choi et al., 2017)" ], [ "Model", "SSE (Nie and Bansal, 2017)" ], [ "Model", "ReSan (Shen et al., 2018)" ], [ "Model", "GloVe BiLSTM-Max (8.6M)" ], [ "Model", "FastText BiLSTM-Max (8.6M)" ], [ "Model", "Naive baseline (9.8M)" ], [ "Model", "Naive baseline (61.3M)" ], [ "Model", "Unweighted DME (8.6M)" ], [ "Model", "DME (8.6M)" ], [ "Model", "CDME (8.6M)" ], [ "Model", "DME* (9.0M)" ], [ "Model", "CDME* (9.0M)" ] ], "column_header_level": 1, "column_headers": [ [ "SNLI" ], [ "MNLI" ] ], "contents": [ [ "84.5", "-" ], [ "84.6", "-" ], [ "86.0", "-" ], [ "86.1", "73.6" ], [ "86.3", "-" ], [ "85.2\u0081}.3", "70.0\u0081}.5" ], [ "85.2\u0081}.2", "70.3\u0081}.3" ], [ "85.6\u0081}.3", "71.1\u0081}.2" ], [ "86.0\u0081}.5", "73.0\u0081}.2" ], [ "86.3\u0081}.4", "74.4\u0081}.2" ], [ "86.2\u0081}.2", "74.4\u0081}.2" ], [ "86.4\u0081}.3", "74.1\u0081}.2" ], [ "86.7\u0081}.2", "74.3\u0081}.4" ], [ "86.5\u0081}.2", "74.9\u0081}.5" ] ], "metrics_loc": "column", "metrics_type": "['Accuracy', 'Accuracy']", "target_entity": "['Unweighted DME (8.6M)', 'DME (8.6M)', 'CDME (8.6M)', 'DME* (9.0M)', 'CDME* (9.0M)']", "table_html": "
SNLI MNLI
Model || InferSent (Conneau et al., 2017) 84.5 -
Model || NSE (Munkhdalai and Yu, 2017) 84.6 -
Model || G-TreeLSTM (Choi et al., 2017) 86.0 -
Model || SSE (Nie and Bansal, 2017) 86.1 73.6
Model || ReSan (Shen et al., 2018) 86.3 -
Model || GloVe BiLSTM-Max (8.6M) 85.2\u0081}.3 70.0\u0081}.5
Model || FastText BiLSTM-Max (8.6M) 85.2\u0081}.2 70.3\u0081}.3
Model || Naive baseline (9.8M) 85.6\u0081}.3 71.1\u0081}.2
Model || Naive baseline (61.3M) 86.0\u0081}.5 73.0\u0081}.2
Model || Unweighted DME (8.6M) 86.3\u0081}.4 74.4\u0081}.2
Model || DME (8.6M) 86.2\u0081}.2 74.4\u0081}.2
Model || CDME (8.6M) 86.4\u0081}.3 74.1\u0081}.2
Model || DME* (9.0M) 86.7\u0081}.2 74.3\u0081}.4
Model || CDME* (9.0M) 86.5\u0081}.2 74.9\u0081}.5
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "D18-1176", "page_no": 5, "dir": "emnlp2018", "description": "4.2 Results. Table 1 shows the results. We report accuracy scores averaged over five runs with different random seeds, together with their standard deviation, for the SNLI and MultiNLI datasets. We include two versions of the naive baseline: one with a 512dimensional BiLSTM encoder; and a bigger one with 2048 dimensions. Both naive baseline models outperform the single encoders that have only GloVe or FastText embeddings. This shows how including more than one embeddings can help performance. Next, we observe that the DME embeddings outperform the naive concatenation baselines, while having fewer parameters. Differences between the three DME variants are small and not significant, although we do note that we found the highest maximum performance for the contextualized version, which adds very few additional parameters. It is important to note that the imposition of weighting thus is not detrimental to performance, which means that DME and CDME provide additional interpretability without sacrificing performance. Finally, we obtain results for using the six different embedding types (marked *), and show that adding in more embeddings increases performance further. To our knowledge, these numbers constitute the state of the art within the model class of single sentence encoders on these tasks.", "class_sentence": "[2, 1, 1, 2, 1, 2, 1, 1, 2, 1, 2]", "sentences": "['4.2 Results.', 'Table 1 shows the results.', 'We report accuracy scores averaged over five runs with different random seeds, together with their standard deviation, for the SNLI and MultiNLI datasets.', 'We include two versions of the naive baseline: one with a 512dimensional BiLSTM encoder; and a bigger one with 2048 dimensions.', 'Both naive baseline models outperform the single encoders that have only GloVe or FastText embeddings.', 'This shows how including more than one embeddings can help performance.', 'Next, we observe that the DME embeddings outperform the naive concatenation baselines, while having fewer parameters.', 'Differences between the three DME variants are small and not significant, although we do note that we found the highest maximum performance for the contextualized version, which adds very few additional parameters.', 'It is important to note that the imposition of weighting thus is not detrimental to performance, which means that DME and CDME provide additional interpretability without sacrificing performance.', 'Finally, we obtain results for using the six different embedding types (marked *), and show that adding in more embeddings increases performance further.', 'To our knowledge, these numbers constitute the state of the art within the model class of single sentence encoders on these tasks.']", "header_mention": "[None, None, ['SNLI', 'MNLI'], ['Naive baseline (9.8M)', 'Naive baseline (61.3M)'], ['Naive baseline (9.8M)', 'Naive baseline (61.3M)', 'GloVe BiLSTM-Max (8.6M)', 'FastText BiLSTM-Max (8.6M)'], None, ['DME (8.6M)', 'Naive baseline (9.8M)', 'Naive baseline (61.3M)'], ['Unweighted DME (8.6M)', 'DME (8.6M)', 'CDME (8.6M)'], ['DME (8.6M)', 'CDME (8.6M)'], ['DME* (9.0M)', 'CDME* (9.0M)'], None]", "valid": 1, "image_id": "D18-1176-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Accuracy scores on the Stanford Natural Language Inference (SNLI) and MultiNLI Mismatched (MNLI) tasks. DME=Dynamic Meta-Embeddings; CDME=Contextualized Dynamic Meta-Embeddings; *=multiple different embedding sets (see Section 4). Number of parameters included in parenthesis. Results averaged over five runs with different random seeds, using a BiLSTM-Max sentence encoder.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n SNLIMNLI
Model || InferSent (Conneau et al., 2017)84.5-
Model || NSE (Munkhdalai and Yu, 2017)84.6-
Model || G-TreeLSTM (Choi et al., 2017)86.0-
Model || SSE (Nie and Bansal, 2017)86.173.6
Model || ReSan (Shen et al., 2018)86.3-
Model || GloVe BiLSTM-Max (8.6M)85.2\u0081}.370.0\u0081}.5
Model || FastText BiLSTM-Max (8.6M)85.2\u0081}.270.3\u0081}.3
Model || Naive baseline (9.8M)85.6\u0081}.371.1\u0081}.2
Model || Naive baseline (61.3M)86.0\u0081}.573.0\u0081}.2
Model || Unweighted DME (8.6M)86.3\u0081}.474.4\u0081}.2
Model || DME (8.6M)86.2\u0081}.274.4\u0081}.2
Model || CDME (8.6M)86.4\u0081}.374.1\u0081}.2
Model || DME* (9.0M)86.7\u0081}.274.3\u0081}.4
Model || CDME* (9.0M)86.5\u0081}.274.9\u0081}.5
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Accuracy scores on the Stanford Natural Language Inference (SNLI) and MultiNLI Mismatched (MNLI) tasks. DME$=$Dynamic Meta-Embeddings; CDME$=$Contextualized Dynamic Meta-Embeddings; *$=$multiple different embedding sets (see Section 4). Number of parameters included in parenthesis. Results averaged over five runs with different random seeds, using a BiLSTM-Max sentence encoder.}\n\\begin{tabular}{lll}\n\\toprule\n & SNLI & MNLI \\\\\n\\midrule\nModel $\\parallel$ InferSent (Conneau et al., 2017) & 84.5 & - \\\\\nModel $\\parallel$ NSE (Munkhdalai and Yu, 2017) & 84.6 & - \\\\\nModel $\\parallel$ G-TreeLSTM (Choi et al., 2017) & 86.0 & - \\\\\nModel $\\parallel$ SSE (Nie and Bansal, 2017) & 86.1 & 73.6 \\\\\nModel $\\parallel$ ReSan (Shen et al., 2018) & 86.3 & - \\\\\nModel $\\parallel$ GloVe BiLSTM-Max (8.6M) & 85.2$\\pm$.3 & 70.0$\\pm$.5 \\\\\nModel $\\parallel$ FastText BiLSTM-Max (8.6M) & 85.2$\\pm$.2 & 70.3$\\pm$.3 \\\\\nModel $\\parallel$ Naive baseline (9.8M) & 85.6$\\pm$.3 & 71.1$\\pm$.2 \\\\\nModel $\\parallel$ Naive baseline (61.3M) & 86.0$\\pm$.5 & 73.0$\\pm$.2 \\\\\nModel $\\parallel$ Unweighted DME (8.6M) & 86.3$\\pm$.4 & 74.4$\\pm$.2 \\\\\nModel $\\parallel$ DME (8.6M) & 86.2$\\pm$.2 & 74.4$\\pm$.2 \\\\\nModel $\\parallel$ CDME (8.6M) & 86.4$\\pm$.3 & 74.1$\\pm$.2 \\\\\nModel $\\parallel$ DME* (9.0M) & 86.7$\\pm$.2 & 74.3$\\pm$.4 \\\\\nModel $\\parallel$ CDME* (9.0M) & 86.5$\\pm$.2 & 74.9$\\pm$.5 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Accuracy scores on the Stanford Natural Language Inference (SNLI) and MultiNLI Mismatched (MNLI) tasks. DME=Dynamic Meta-Embeddings; CDME=Contextualized Dynamic Meta-Embeddings; *=multiple different embedding sets (see Section 4). Number of parameters included in parenthesis. Results averaged over five runs with different random seeds, using a BiLSTM-Max sentence encoder.\n
\n \n SNLI\n \n MNLI\n
\n Model || InferSent (Conneau et al., 2017)\n \n 84.5\n \n -\n
\n Model || NSE (Munkhdalai and Yu, 2017)\n \n 84.6\n \n -\n
\n Model || G-TreeLSTM (Choi et al., 2017)\n \n 86.0\n \n -\n
\n Model || SSE (Nie and Bansal, 2017)\n \n 86.1\n \n 73.6\n
\n Model || ReSan (Shen et al., 2018)\n \n 86.3\n \n -\n
\n Model || GloVe BiLSTM-Max (8.6M)\n \n 85.2\u0081}.3\n \n 70.0\u0081}.5\n
\n Model || FastText BiLSTM-Max (8.6M)\n \n 85.2\u0081}.2\n \n 70.3\u0081}.3\n
\n Model || Naive baseline (9.8M)\n \n 85.6\u0081}.3\n \n 71.1\u0081}.2\n
\n Model || Naive baseline (61.3M)\n \n 86.0\u0081}.5\n \n 73.0\u0081}.2\n
\n Model || Unweighted DME (8.6M)\n \n 86.3\u0081}.4\n \n 74.4\u0081}.2\n
\n Model || DME (8.6M)\n \n 86.2\u0081}.2\n \n 74.4\u0081}.2\n
\n Model || CDME (8.6M)\n \n 86.4\u0081}.3\n \n 74.1\u0081}.2\n
\n Model || DME* (9.0M)\n \n 86.7\u0081}.2\n \n 74.3\u0081}.4\n
\n Model || CDME* (9.0M)\n \n 86.5\u0081}.2\n \n 74.9\u0081}.5\n
\n" }, { "instance_id": "3d372d68-b0a9-4a52-895e-bfd473d1b1ca", "table_id_paper": "D18-1188table_2", "caption": "Performance of different approaches on the WikiSQL dataset. The two evaluation metrics are logical form accuracy (Acclf ) and execution accuracy (Accex). The settings of the training data represent the proportion of supervised data we use.", "row_header_level": 4, "row_headers": [ [ "Methods", "Attentional Seq2Seq", "Training Data", "100%" ], [ "Methods", "Aug.PntNet (Zhong et al. 2017)", "Training Data", "100%" ], [ "Methods", "Aug.PntNet (re-implemented by us)", "Training Data", "100%" ], [ "Methods", "Seq2SQL (Zhong et al. 2017)", "Training Data", "100%" ], [ "Methods", "SQLNet (Xu et al. 2017)", "Training Data", "100%" ], [ "Methods", "STAMP", "Training Data", "30%" ], [ "Methods", "STAMP + QG", "Training Data", "30%" ], [ "Methods", "STAMP", "Training Data", "100%" ], [ "Methods", "STAMP + QG", "Training Data", "100%" ] ], "column_header_level": 2, "column_headers": [ [ "Dev", "Acclf" ], [ "Dev", "Accex" ], [ "Test", "Acclf" ], [ "Test", "Accex" ] ], "contents": [ [ "23.3%", "37.0%", "23.4%", "35.9%" ], [ "44.1%", "53.8%", "43.3%", "53.3%" ], [ "51.5%", "58.9%", "52.1%", "59.2%" ], [ "49.5%", "60.8%", "48.3%", "59.4%" ], [ "-", "69.8%", "-", "68.0%" ], [ "54.6%", "69.7%", "53.7%", "68.9%" ], [ "61.6%", "74.4%", "61.2%", "73.9%" ], [ "61.5%", "74.8%", "60.7%", "74.4%" ], [ "64.3%", "76.5%", "63.7%", "75.5%" ] ], "metrics_loc": "column", "metrics_type": "['Acclf', 'Accex', 'Acclf', 'Accex']", "target_entity": "['STAMP', 'STAMP + QG']", "table_html": "
Dev || Acclf Dev || Accex Test || Acclf Test || Accex
Methods || Attentional Seq2Seq || Training Data || 100% 23.3% 37.0% 23.4% 35.9%
Methods || Aug.PntNet (Zhong et al. 2017) || Training Data || 100% 44.1% 53.8% 43.3% 53.3%
Methods || Aug.PntNet (re-implemented by us) || Training Data || 100% 51.5% 58.9% 52.1% 59.2%
Methods || Seq2SQL (Zhong et al. 2017) || Training Data || 100% 49.5% 60.8% 48.3% 59.4%
Methods || SQLNet (Xu et al. 2017) || Training Data || 100% - 69.8% - 68.0%
Methods || STAMP || Training Data || 30% 54.6% 69.7% 53.7% 68.9%
Methods || STAMP + QG || Training Data || 30% 61.6% 74.4% 61.2% 73.9%
Methods || STAMP || Training Data || 100% 61.5% 74.8% 60.7% 74.4%
Methods || STAMP + QG || Training Data || 100% 64.3% 76.5% 63.7% 75.5%
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1188", "page_no": 5, "dir": "emnlp2018", "description": "From Table 2, we can see that STAMP performs better than existing systems when trained on the full WikiSQL training dataset, achieving state-ofthe-art execution accuracy and logical form accuracy on WikiSQL. We further conduct experiments to demonstrate the effectiveness of our question generation driven approach. We run the entire pipeline (STAMP+QG) with different percentages of training data. The second column \u201cTraining Data\u201d in Table 2 and the x-axis in Figure 3 represent the proportion of WikiSQL training data we use for training the QG model and semantic parser. That is to say, STAMP +QG with 30% means that we sample 30% WikiSQL training data to train the QG model, and then combine QG generated data and exactly the same 30% WikiSQL training data we sampled before to train the semantic parser. In this experiment, we sample five SQL queries for each table in the training data, resulting in 43.5K SQL queries. Applying the QG model on these SQL queries, we get 92.8K SQLquestion pairs. From Figure 3, we see that accuracy increases as the amount of supervised training data expands. Results show that QG empowers the STAMP model to achieve the same accuracy on WikiSQL dataset with 30% of the training data. Applying QG to the STAMP model under the full setting brings further improvements, resulting in new state-of-the-art accuracies.", "class_sentence": "[1, 2, 2, 1, 2, 2, 2, 0, 1, 1]", "sentences": "['From Table 2, we can see that STAMP performs better than existing systems when trained on the full WikiSQL training dataset, achieving state-ofthe-art execution accuracy and logical form accuracy on WikiSQL.', 'We further conduct experiments to demonstrate the effectiveness of our question generation driven approach.', 'We run the entire pipeline (STAMP+QG) with different percentages of training data.', 'The second column \u201cTraining Data\u201d in Table 2 and the x-axis in Figure 3 represent the proportion of WikiSQL training data we use for training the QG model and semantic parser.', 'That is to say, STAMP +QG with 30% means that we sample 30% WikiSQL training data to train the QG model, and then combine QG generated data and exactly the same 30% WikiSQL training data we sampled before to train the semantic parser.', 'In this experiment, we sample five SQL queries for each table in the training data, resulting in 43.5K SQL queries.', 'Applying the QG model on these SQL queries, we get 92.8K SQLquestion pairs.', 'From Figure 3, we see that accuracy increases as the amount of supervised training data expands.', 'Results show that QG empowers the STAMP model to achieve the same accuracy on WikiSQL dataset with 30% of the training data.', 'Applying QG to the STAMP model under the full setting brings further improvements, resulting in new state-of-the-art accuracies.']", "header_mention": "[['STAMP', '100%', 'Acclf', 'Accex'], None, ['STAMP + QG', '30%', '100%'], ['Training Data'], ['STAMP + QG', '30%'], None, ['STAMP + QG'], None, ['STAMP + QG', '30%'], ['STAMP + QG', '100%', 'Attentional Seq2Seq', 'Aug.PntNet (Zhong et al. 2017)', 'Aug.PntNet (re-implemented by us)', 'Seq2SQL (Zhong et al. 2017)', 'SQLNet (Xu et al. 2017)']]", "valid": 1, "image_id": "D18-1188-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance of different approaches on the WikiSQL dataset. The two evaluation metrics are logical form accuracy (Acclf ) and execution accuracy (Accex). The settings of the training data represent the proportion of supervised data we use.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Dev || AcclfDev || AccexTest || AcclfTest || Accex
Methods || Attentional Seq2Seq || Training Data || 100%23.3%37.0%23.4%35.9%
Methods || Aug.PntNet (Zhong et al. 2017) || Training Data || 100%44.1%53.8%43.3%53.3%
Methods || Aug.PntNet (re-implemented by us) || Training Data || 100%51.5%58.9%52.1%59.2%
Methods || Seq2SQL (Zhong et al. 2017) || Training Data || 100%49.5%60.8%48.3%59.4%
Methods || SQLNet (Xu et al. 2017) || Training Data || 100%-69.8%-68.0%
Methods || STAMP || Training Data || 30%54.6%69.7%53.7%68.9%
Methods || STAMP + QG || Training Data || 30%61.6%74.4%61.2%73.9%
Methods || STAMP || Training Data || 100%61.5%74.8%60.7%74.4%
Methods || STAMP + QG || Training Data || 100%64.3%76.5%63.7%75.5%
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance of different approaches on the WikiSQL dataset. The two evaluation metrics are logical form accuracy (Acclf ) and execution accuracy (Accex). The settings of the training data represent the proportion of supervised data we use.}\n\\begin{tabular}{lllll}\n\\toprule\n & Dev $\\parallel$ Acclf & Dev $\\parallel$ Accex & Test $\\parallel$ Acclf & Test $\\parallel$ Accex \\\\\n\\midrule\nMethods $\\parallel$ Attentional Seq2Seq $\\parallel$ Training Data $\\parallel$ 100\\% & 23.3\\% & 37.0\\% & 23.4\\% & 35.9\\% \\\\\nMethods $\\parallel$ Aug.PntNet (Zhong et al. 2017) $\\parallel$ Training Data $\\parallel$ 100\\% & 44.1\\% & 53.8\\% & 43.3\\% & 53.3\\% \\\\\nMethods $\\parallel$ Aug.PntNet (re-implemented by us) $\\parallel$ Training Data $\\parallel$ 100\\% & 51.5\\% & 58.9\\% & 52.1\\% & 59.2\\% \\\\\nMethods $\\parallel$ Seq2SQL (Zhong et al. 2017) $\\parallel$ Training Data $\\parallel$ 100\\% & 49.5\\% & 60.8\\% & 48.3\\% & 59.4\\% \\\\\nMethods $\\parallel$ SQLNet (Xu et al. 2017) $\\parallel$ Training Data $\\parallel$ 100\\% & - & 69.8\\% & - & 68.0\\% \\\\\nMethods $\\parallel$ STAMP $\\parallel$ Training Data $\\parallel$ 30\\% & 54.6\\% & 69.7\\% & 53.7\\% & 68.9\\% \\\\\nMethods $\\parallel$ STAMP + QG $\\parallel$ Training Data $\\parallel$ 30\\% & 61.6\\% & 74.4\\% & 61.2\\% & 73.9\\% \\\\\nMethods $\\parallel$ STAMP $\\parallel$ Training Data $\\parallel$ 100\\% & 61.5\\% & 74.8\\% & 60.7\\% & 74.4\\% \\\\\nMethods $\\parallel$ STAMP + QG $\\parallel$ Training Data $\\parallel$ 100\\% & 64.3\\% & 76.5\\% & 63.7\\% & 75.5\\% \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Performance of different approaches on the WikiSQL dataset. The two evaluation metrics are logical form accuracy (Acclf ) and execution accuracy (Accex). The settings of the training data represent the proportion of supervised data we use.\n
\n \n Dev || Acclf\n \n Dev || Accex\n \n Test || Acclf\n \n Test || Accex\n
\n Methods || Attentional Seq2Seq || Training Data || 100%\n \n 23.3%\n \n 37.0%\n \n 23.4%\n \n 35.9%\n
\n Methods || Aug.PntNet (Zhong et al. 2017) || Training Data || 100%\n \n 44.1%\n \n 53.8%\n \n 43.3%\n \n 53.3%\n
\n Methods || Aug.PntNet (re-implemented by us) || Training Data || 100%\n \n 51.5%\n \n 58.9%\n \n 52.1%\n \n 59.2%\n
\n Methods || Seq2SQL (Zhong et al. 2017) || Training Data || 100%\n \n 49.5%\n \n 60.8%\n \n 48.3%\n \n 59.4%\n
\n Methods || SQLNet (Xu et al. 2017) || Training Data || 100%\n \n -\n \n 69.8%\n \n -\n \n 68.0%\n
\n Methods || STAMP || Training Data || 30%\n \n 54.6%\n \n 69.7%\n \n 53.7%\n \n 68.9%\n
\n Methods || STAMP + QG || Training Data || 30%\n \n 61.6%\n \n 74.4%\n \n 61.2%\n \n 73.9%\n
\n Methods || STAMP || Training Data || 100%\n \n 61.5%\n \n 74.8%\n \n 60.7%\n \n 74.4%\n
\n Methods || STAMP + QG || Training Data || 100%\n \n 64.3%\n \n 76.5%\n \n 63.7%\n \n 75.5%\n
\n" }, { "instance_id": "55715140-30f8-42b9-a8ae-5b03f0880836", "table_id_paper": "D18-1219table_5", "caption": "Results of using NP head alone for bridging anaphora resolution based on different word representation resources. Bold indicates statistically significant differences over the baselines (two-sided paired approximate randomization test, p < 0.01).", "row_header_level": 1, "row_headers": [ [ "GloVe GigaWiki14" ], [ "GloVe Giga" ], [ "embeddings PP" ], [ "embeddings bridging" ] ], "column_header_level": 1, "column_headers": [ [ "acc" ] ], "contents": [ [ "21.42" ], [ "21.87" ], [ "33.03" ], [ "34.84" ] ], "metrics_loc": "column", "metrics_type": "['acc']", "target_entity": "['embeddings bridging']", "table_html": "
acc
GloVe GigaWiki14 21.42
GloVe Giga 21.87
embeddings PP 33.03
embeddings bridging 34.84
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "D18-1219", "page_no": 7, "dir": "emnlp2018", "description": "Table 5 lists the results for bridging anaphora resolution based on different word representation resources. We notice that there is not much difference between GloVe GigaWiki14 and GloVe Giga. We find that using embeddings PP achieves an accuracy of 33.03% on the ISNotes corpus, which outperforms the results based on GloVe GigaWiki14 and GloVe Giga by a large margin. Using embeddings bridging further improves the result by 1.8%. Although the improvement is not significant, we suspect that the representations for words without the suffix \u201c PP\u201d in embeddings bridging are more accurate because they are trained on many more instances in the vanilla GloVe vectors (GloVe GigaWiki14).", "class_sentence": "[1, 1, 1, 1, 2]", "sentences": "['Table 5 lists the results for bridging anaphora resolution based on different word representation resources.', 'We notice that there is not much difference between GloVe GigaWiki14 and GloVe Giga.', 'We find that using embeddings PP achieves an accuracy of 33.03% on the ISNotes corpus, which outperforms the results based on GloVe GigaWiki14 and GloVe Giga by a large margin.', 'Using embeddings bridging further improves the result by 1.8%.', 'Although the improvement is not significant, we suspect that the representations for words without the suffix \u201c PP\u201d in embeddings bridging are more accurate because they are trained on many more instances in the vanilla GloVe vectors (GloVe GigaWiki14).']", "header_mention": "[['GloVe GigaWiki14', 'GloVe Giga', 'embeddings PP', 'embeddings bridging'], ['GloVe GigaWiki14', 'GloVe Giga'], ['embeddings PP', 'GloVe GigaWiki14', 'GloVe Giga', 'acc'], ['embeddings bridging', 'embeddings PP', 'acc'], ['GloVe GigaWiki14']]", "valid": 1, "image_id": "D18-1219-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results of using NP head alone for bridging anaphora resolution based on different word representation resources. Bold indicates statistically significant differences over the baselines (two-sided paired approximate randomization test, p < 0.01).

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n acc
GloVe GigaWiki1421.42
GloVe Giga21.87
embeddings PP33.03
embeddings bridging34.84
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results of using NP head alone for bridging anaphora resolution based on different word representation resources. Bold indicates statistically significant differences over the baselines (two-sided paired approximate randomization test, p $<$ 0.01).}\n\\begin{tabular}{lr}\n\\toprule\n & acc \\\\\n\\midrule\nGloVe GigaWiki14 & 21.42 \\\\\nGloVe Giga & 21.87 \\\\\nembeddings PP & 33.03 \\\\\nembeddings bridging & 34.84 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Results of using NP head alone for bridging anaphora resolution based on different word representation resources. Bold indicates statistically significant differences over the baselines (two-sided paired approximate randomization test, p < 0.01).\n
\n \n acc\n
\n GloVe GigaWiki14\n \n 21.42\n
\n GloVe Giga\n \n 21.87\n
\n embeddings PP\n \n 33.03\n
\n embeddings bridging\n \n 34.84\n
\n" }, { "instance_id": "5994f5b0-1a80-4d89-a0bf-fb89d1ad2c78", "table_id_paper": "D18-1262table_5", "caption": "Comparison of our Syn-GCN model with (Marcheggiani and Titov, 2017), (He et al., 2018) and (Cai et al., 2018) on the English test set. \u2206 F1 shows the absolute performance gap between syntax-agnostic and syntax-aware settings.", "row_header_level": 2, "row_headers": [ [ "System", "M&T (2017)" ], [ "System", "He et al. (2018)" ], [ "System", "Cai et al. (2018)" ], [ "System", "Our model" ] ], "column_header_level": 1, "column_headers": [ [ "syntax-agnostic" ], [ "syntax-aware" ], [ "\u0394F1" ] ], "contents": [ [ "87.7", "88.0", "0.3" ], [ "88.7", "89.5", "0.8" ], [ "89.6", "89.6", "\u22480.0" ], [ "88.7", "89.8", "1.1" ] ], "metrics_loc": "column", "metrics_type": "['F1', 'F1', 'F1']", "target_entity": "['Our model']", "table_html": "
syntax-agnostic syntax-aware \u0394F1
System || M&T (2017) 87.7 88.0 0.3
System || He et al. (2018) 88.7 89.5 0.8
System || Cai et al. (2018) 89.6 89.6 \u22480.0
System || Our model 88.7 89.8 1.1
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "D18-1262", "page_no": 7, "dir": "emnlp2018", "description": "Effect of deep encoder. Table 5 reports F1 scores of our Syn-GCN model, Marcheggiani and Titov (2017), He et al. (2018) and Cai et al. (2018) on English test set in both syntax-agnostic and syntax-aware settings. The comparison shows that our framework is more effective for incorporating syntactic information by giving more performance improvement through introducing syntax over syntax-agnostic SRL than previous state-ofthe-art systems did.", "class_sentence": "[2, 1, 1]", "sentences": "['Effect of deep encoder.', 'Table 5 reports F1 scores of our Syn-GCN model, Marcheggiani and Titov (2017), He et al. (2018) and Cai et al. (2018) on English test set in both syntax-agnostic and syntax-aware settings.', 'The comparison shows that our framework is more effective for incorporating syntactic information by giving more performance improvement through introducing syntax over syntax-agnostic SRL than previous state-ofthe-art systems did.']", "header_mention": "[None, ['M&T (2017)', 'He et al. (2018)', 'Cai et al. (2018)', 'Our model', 'syntax-agnostic', 'syntax-aware'], ['M&T (2017)', 'He et al. (2018)', 'Cai et al. (2018)', 'Our model']]", "valid": 1, "image_id": "D18-1262-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison of our Syn-GCN model with (Marcheggiani and Titov, 2017), (He et al., 2018) and (Cai et al., 2018) on the English test set. \u2206 F1 shows the absolute performance gap between syntax-agnostic and syntax-aware settings.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n syntax-agnosticsyntax-aware\u0394F1
System || M&T (2017)87.788.00.3
System || He et al. (2018)88.789.50.8
System || Cai et al. (2018)89.689.6\u22480.0
System || Our model88.789.81.1
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison of our Syn-GCN model with (Marcheggiani and Titov, 2017), (He et al., 2018) and (Cai et al., 2018) on the English test set. $\\Delta$ F1 shows the absolute performance gap between syntax-agnostic and syntax-aware settings.}\n\\begin{tabular}{lrrl}\n\\toprule\n & syntax-agnostic & syntax-aware & $\\Delta$F1 \\\\\n\\midrule\nSystem $\\parallel$ M\\&T (2017) & 87.7 & 88 & 0.3 \\\\\nSystem $\\parallel$ He et al. (2018) & 88.7 & 89.5 & 0.8 \\\\\nSystem $\\parallel$ Cai et al. (2018) & 89.6 & 89.6 & $\\approx$0.0 \\\\\nSystem $\\parallel$ Our model & 88.7 & 89.8 & 1.1 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Comparison of our Syn-GCN model with (Marcheggiani and Titov, 2017), (He et al., 2018) and (Cai et al., 2018) on the English test set. \u2206 F1 shows the absolute performance gap between syntax-agnostic and syntax-aware settings.\n
\n \n syntax-agnostic\n \n syntax-aware\n \n \u0394F1\n
\n System || M&T (2017)\n \n 87.7\n \n 88.0\n \n 0.3\n
\n System || He et al. (2018)\n \n 88.7\n \n 89.5\n \n 0.8\n
\n System || Cai et al. (2018)\n \n 89.6\n \n 89.6\n \n \u22480.0\n
\n System || Our model\n \n 88.7\n \n 89.8\n \n 1.1\n
\n" }, { "instance_id": "ef09ad48-da3c-49e5-acf7-2cf9c756c2f6", "table_id_paper": "D18-1273table_6", "caption": "Human evaluation results. S-ocr refer to 150 sentences from the D-ocr and E-ocr represents the errors in S-ocr, similar for S-asr and Easr. R denotes the average recall of three students. Numbers in bold denotes the correctly-annotated results by students.", "row_header_level": 1, "row_headers": [ [ "S-ocr" ], [ "E-ocr" ], [ "S-asr" ], [ "E-asr" ] ], "column_header_level": 1, "column_headers": [ [ "Correctly-annotated Number of Stu1" ], [ "Correctly-annotated Number of Stu2" ], [ "Correctly-annotated Number of Stu3" ], [ "R" ] ], "contents": [ [ "84/150", "100/150", "75/150", "57.3" ], [ "104/170", "121/170", "100/170", "72.0" ], [ "95/150", "79/150", "106/150", "62.0" ], [ "341/393", "179/393", "356/393", "74.3" ] ], "metrics_loc": "column", "metrics_type": "['Correctly-annotated Number of Stu1', 'Correctly-annotated Number of Stu2', 'Correctly-annotated Number of Stu3', 'R']", "target_entity": "['S-ocr', 'E-ocr', 'S-asr', 'E-asr']", "table_html": "
Stu1 Stu2 Stu3 R
S-ocr 84/150 100/150 75/150 57.3
E-ocr 104/170 121/170 100/170 72.0
S-asr 95/150 79/150 106/150 62.0
E-asr 341/393 179/393 356/393 74.3
", "table_name": "Table 6", "table_id": "table_6", "paper_id": "D18-1273", "page_no": 6, "dir": "emnlp2018", "description": "Results. Table 6 shows the information of 300 sentences and the annotation results on them. The average recall in the table illustrates that three students have a higher recognition rate for errors from D-asr than that from D-ocr, which, to some extent, indicates that P-style errors are easier to be detected than V-style ones. Besides, we observe that three volunteers fail to identify around 36.9% errors on average, which may indicate that our generated sentences include some challenging errors which are likely to be made by human. Such errors are valuable for CSC since they are potential real cases in people\u00e2\u20ac\u2122s writing or typing.", "class_sentence": "[2, 1, 1, 1, 2]", "sentences": "['Results.', 'Table 6 shows the information of 300 sentences and the annotation results on them.', 'The average recall in the table illustrates that three students have a higher recognition rate for errors from D-asr than that from D-ocr, which, to some extent, indicates that P-style errors are easier to be detected than V-style ones.', 'Besides, we observe that three volunteers fail to identify around 36.9% errors on average, which may indicate that our generated sentences include some challenging errors which are likely to be made by human.', 'Such errors are valuable for CSC since they are potential real cases in people\u00e2\u20ac\u2122s writing or typing.']", "header_mention": "[None, None, ['E-asr', 'E-ocr', 'R'], ['E-asr', 'E-ocr', 'R'], None]", "valid": 1, "image_id": "D18-1273-Table6-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Human evaluation results. S-ocr refer to 150 sentences from the D-ocr and E-ocr represents the errors in S-ocr, similar for S-asr and Easr. R denotes the average recall of three students. Numbers in bold denotes the correctly-annotated results by students.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Stu1Stu2Stu3R
S-ocr84/150100/15075/15057.3
E-ocr104/170121/170100/17072.0
S-asr95/15079/150106/15062.0
E-asr341/393179/393356/39374.3
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{5}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Human evaluation results. S-ocr refer to 150 sentences from the D-ocr and E-ocr represents the errors in S-ocr, similar for S-asr and Easr. R denotes the average recall of three students. Numbers in bold denotes the correctly-annotated results by students.}\n\\begin{tabular}{llllr}\n\\toprule\n & Stu1 & Stu2 & Stu3 & R \\\\\n\\midrule\nS-ocr & 84/150 & 100/150 & 75/150 & 57.3 \\\\\nE-ocr & 104/170 & 121/170 & 100/170 & 72 \\\\\nS-asr & 95/150 & 79/150 & 106/150 & 62 \\\\\nE-asr & 341/393 & 179/393 & 356/393 & 74.3 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 6: Human evaluation results. S-ocr refer to 150 sentences from the D-ocr and E-ocr represents the errors in S-ocr, similar for S-asr and Easr. R denotes the average recall of three students. Numbers in bold denotes the correctly-annotated results by students.\n
\n \n Stu1\n \n Stu2\n \n Stu3\n \n R\n
\n S-ocr\n \n 84/150\n \n 100/150\n \n 75/150\n \n 57.3\n
\n E-ocr\n \n 104/170\n \n 121/170\n \n 100/170\n \n 72.0\n
\n S-asr\n \n 95/150\n \n 79/150\n \n 106/150\n \n 62.0\n
\n E-asr\n \n 341/393\n \n 179/393\n \n 356/393\n \n 74.3\n
\n" }, { "instance_id": "300649df-960f-4108-9711-9b65cb7f4598", "table_id_paper": "D18-1276table_2", "caption": "Performance of Clique-EBM with pruned edges in G.", "row_header_level": 2, "row_headers": [ [ "k", "5" ], [ "k", "10" ], [ "k", "15" ], [ "k", "20" ] ], "column_header_level": 2, "column_headers": [ [ "WPT", "P" ], [ "WPT", "R" ], [ "WPT", "F" ], [ "WP3T", "P" ], [ "WP3T", "R" ], [ "WP3T", "F" ] ], "contents": [ [ "90.46", "92.27", "91.36", "83.52", "80.48", "81.97" ], [ "92.92", "95.07", "93.98", "85.32", "84.4", "84.86" ], [ "94.85", "96.14", "95.49", "87.67", "86.38", "87.02" ], [ "95.23", "96.49", "95.86", "89.25", "88.62", "88.93" ] ], "metrics_loc": "column", "metrics_type": "['P', 'R', 'F', 'P', 'R', 'F']", "target_entity": "['k']", "table_html": "
WPT || P WPT || R WPT || F WP3T || P WP3T || R WP3T || F
k || 5 90.46 92.27 91.36 83.52 80.48 81.97
k || 10 92.92 95.07 93.98 85.32 84.4 84.86
k || 15 94.85 96.14 95.49 87.67 86.38 87.02
k || 20 95.23 96.49 95.86 89.25 88.62 88.93
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1276", "page_no": 8, "dir": "emnlp2018", "description": "For any two words appearing in an exhaustive segmentation, we keep an edge only if both the words overlap within a distance of k characters. We experiment with k = 5, 10, 15 and 20. Hence, for K = 20, a word will form edges with all the words that fall within 20 characters to left and 20 characters to right. The average length of an input sequence in DCS10K is 40.79 characters. We do not modify our inference procedure in system 8 other than to take care of the possibility that a clique need not always be returned. Table 2 shows the results for different values of k. Interestingly, the results show a monotonic increase with the increase in context window, and the results with the entire context are still better than those with k = 20, even though only marginally. It is interesting to note that, keeping the entire context does not adversely affect the predictions as none of the pruned models outperforms System 8.", "class_sentence": "[2, 2, 2, 2, 2, 1, 1, 2]", "sentences": "['For any two words appearing in an exhaustive segmentation, we keep an edge only if both the words overlap within a distance of k characters.', 'We experiment with k = 5, 10, 15 and 20.', 'Hence, for K = 20, a word will form edges with all the words that fall within 20 characters to left and 20 characters to right.', 'The average length of an input sequence in DCS10K is 40.79 characters.', 'We do not modify our inference procedure in system 8 other than to take care of the possibility that a clique need not always be returned.', 'Table 2 shows the results for different values of k.', 'Interestingly, the results show a monotonic increase with the increase in context window, and the results with the entire context are still better than those with k = 20, even though only marginally.', 'It is interesting to note that, keeping the entire context does not adversely affect the predictions as none of the pruned models outperforms System 8.']", "header_mention": "[['k'], ['k', '5', '10', '15', '20'], ['k', '20'], None, None, ['k', '5', '10', '15', '20'], ['k', '5', '10', '15', '20'], None]", "valid": 1, "image_id": "D18-1276-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance of Clique-EBM with pruned edges in G.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n WPT || PWPT || RWPT || FWP3T || PWP3T || RWP3T || F
k || 590.4692.2791.3683.5280.4881.97
k || 1092.9295.0793.9885.3284.484.86
k || 1594.8596.1495.4987.6786.3887.02
k || 2095.2396.4995.8689.2588.6288.93
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance of Clique-EBM with pruned edges in G.}\n\\begin{tabular}{lrrrrrr}\n\\toprule\n & WPT $\\parallel$ P & WPT $\\parallel$ R & WPT $\\parallel$ F & WP3T $\\parallel$ P & WP3T $\\parallel$ R & WP3T $\\parallel$ F \\\\\n\\midrule\nk $\\parallel$ 5 & 90.46 & 92.27 & 91.36 & 83.52 & 80.48 & 81.97 \\\\\nk $\\parallel$ 10 & 92.92 & 95.07 & 93.98 & 85.32 & 84.4 & 84.86 \\\\\nk $\\parallel$ 15 & 94.85 & 96.14 & 95.49 & 87.67 & 86.38 & 87.02 \\\\\nk $\\parallel$ 20 & 95.23 & 96.49 & 95.86 & 89.25 & 88.62 & 88.93 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Performance of Clique-EBM with pruned edges in G.\n
\n \n WPT || P\n \n WPT || R\n \n WPT || F\n \n WP3T || P\n \n WP3T || R\n \n WP3T || F\n
\n k || 5\n \n 90.46\n \n 92.27\n \n 91.36\n \n 83.52\n \n 80.48\n \n 81.97\n
\n k || 10\n \n 92.92\n \n 95.07\n \n 93.98\n \n 85.32\n \n 84.4\n \n 84.86\n
\n k || 15\n \n 94.85\n \n 96.14\n \n 95.49\n \n 87.67\n \n 86.38\n \n 87.02\n
\n k || 20\n \n 95.23\n \n 96.49\n \n 95.86\n \n 89.25\n \n 88.62\n \n 88.93\n
\n" }, { "instance_id": "b89b7a9f-f9a1-4388-9d81-44d94ae7a0a7", "table_id_paper": "D18-1281table_2", "caption": "Comparison of various bounding box regressors on Flickr30k Entities for different IoU thresholds. The number of parameters in Gr is also shown.", "row_header_level": 2, "row_headers": [ [ "Architecture", "w/o regression" ], [ "Architecture", "300-16(-4096)" ], [ "Architecture", "300-64(-4096)" ], [ "Architecture", "300-256(-4096)" ], [ "Architecture", "300-1024(-4096)" ], [ "Architecture", "300(-256-4096)" ], [ "Architecture", "300-4096" ] ], "column_header_level": 2, "column_headers": [ [ "Params", "-" ], [ "IoU", "0.5" ], [ "IoU", "0.6" ], [ "IoU", "0.7" ], [ "IoU", "0.8" ], [ "IoU", "0.9" ] ], "contents": [ [ "-", "65.21", "53.19", "35.70", "14.32", "1.88" ], [ "0.3M", "64.14", "57.66", "48.22", "33.04", "9.29" ], [ "1.1M", "63.87", "57.43", "49.05", "33.84", "10.55" ], [ "4.3M", "63.84", "57.70", "48.71", "33.87", "10.05" ], [ "17M", "64.29", "58.05", "48.49", "33.94", "10.09" ], [ "4.5M", "62.82", "56.28", "48.02", "32.71", "9.89" ], [ "1.2M", "63.23", "56.92", "48.17", "32.66", "9.20" ] ], "metrics_loc": "column", "metrics_type": "['# Params', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy']", "target_entity": "['300-16(-4096)', '300-64(-4096)', '300-256(-4096)', '300-1024(-4096)', '300(-256-4096)', '300-4096']", "table_html": "
Params || - IoU || 0.5 IoU || 0.6 IoU || 0.7 IoU || 0.8 IoU || 0.9
Architecture || w/o regression - 65.21 53.19 35.70 14.32 1.88
Architecture || 300-16(-4096) 0.3M 64.14 57.66 48.22 33.04 9.29
Architecture || 300-64(-4096) 1.1M 63.87 57.43 49.05 33.84 10.55
Architecture || 300-256(-4096) 4.3M 63.84 57.70 48.71 33.87 10.05
Architecture || 300-1024(-4096) 17M 64.29 58.05 48.49 33.94 10.09
Architecture || 300(-256-4096) 4.5M 62.82 56.28 48.02 32.71 9.89
Architecture || 300-4096 1.2M 63.23 56.92 48.17 32.66 9.20
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1281", "page_no": 7, "dir": "emnlp2018", "description": "Table 2 shows the results with and without regressor. The regressor significantly improved the accuracy with high IoU thresholds, which demonstrates that the regressor improved the localization accuracy. In addition, the accuracy did not decrease as a result of sharing the hidden layer or reducing the number of units in the hidden layer. This suggests that the regressor lies in a very low-dimensional manifold because the regressor for one concept can be shared by many concepts (e.g., the person regressor can be used for man, woman, girl, boy, etc.). The number of parameters was significantly reduced by these tricks, to even fewer than in the linear transformation. The accuracy slightly decreased with a threshold of 0.5, because the regressor was not learned properly for the categories that did not frequently appear in the training data.", "class_sentence": "[1, 1, 1, 2, 1, 1]", "sentences": "['Table 2 shows the results with and without regressor.', 'The regressor significantly improved the accuracy with high IoU thresholds, which demonstrates that the regressor improved the localization accuracy.', 'In addition, the accuracy did not decrease as a result of sharing the hidden layer or reducing the number of units in the hidden layer.', 'This suggests that the regressor lies in a very low-dimensional manifold because the regressor for one concept can be shared by many concepts (e.g., the person regressor can be used for man, woman, girl, boy, etc.).', 'The number of parameters was significantly reduced by these tricks, to even fewer than in the linear transformation.', 'The accuracy slightly decreased with a threshold of 0.5, because the regressor was not learned properly for the categories that did not frequently appear in the training data.']", "header_mention": "[['w/o regression', '300-16(-4096)', '300-64(-4096)', '300-256(-4096)', '300-1024(-4096)', '300(-256-4096)', '300-4096'], ['300-16(-4096)', '300-64(-4096)', '300-256(-4096)', '300-1024(-4096)', '300(-256-4096)', '300-4096', 'IoU', '0.6', '0.7', '0.8', '0.9'], ['300(-256-4096)', '300-4096'], ['300-16(-4096)', '300-64(-4096)', '300-256(-4096)', '300-1024(-4096)', '300(-256-4096)', '300-4096'], ['Params', '300-4096', '300(-256-4096)'], ['w/o regression', '300-16(-4096)', '300-64(-4096)', '300-256(-4096)', '300-1024(-4096)', '300(-256-4096)', '300-4096', 'IoU', '0.5']]", "valid": 1, "image_id": "D18-1281-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison of various bounding box regressors on Flickr30k Entities for different IoU thresholds. The number of parameters in Gr is also shown.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Params || -IoU || 0.5IoU || 0.6IoU || 0.7IoU || 0.8IoU || 0.9
Architecture || w/o regression-65.2153.1935.7014.321.88
Architecture || 300-16(-4096)0.3M64.1457.6648.2233.049.29
Architecture || 300-64(-4096)1.1M63.8757.4349.0533.8410.55
Architecture || 300-256(-4096)4.3M63.8457.7048.7133.8710.05
Architecture || 300-1024(-4096)17M64.2958.0548.4933.9410.09
Architecture || 300(-256-4096)4.5M62.8256.2848.0232.719.89
Architecture || 300-40961.2M63.2356.9248.1732.669.20
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison of various bounding box regressors on Flickr30k Entities for different IoU thresholds. The number of parameters in Gr is also shown.}\n\\begin{tabular}{llrrrrr}\n\\toprule\n & Params $\\parallel$ - & IoU $\\parallel$ 0.5 & IoU $\\parallel$ 0.6 & IoU $\\parallel$ 0.7 & IoU $\\parallel$ 0.8 & IoU $\\parallel$ 0.9 \\\\\n\\midrule\nArchitecture $\\parallel$ w/o regression & - & 65.21 & 53.19 & 35.7 & 14.32 & 1.88 \\\\\nArchitecture $\\parallel$ 300-16(-4096) & 0.3M & 64.14 & 57.66 & 48.22 & 33.04 & 9.29 \\\\\nArchitecture $\\parallel$ 300-64(-4096) & 1.1M & 63.87 & 57.43 & 49.05 & 33.84 & 10.55 \\\\\nArchitecture $\\parallel$ 300-256(-4096) & 4.3M & 63.84 & 57.7 & 48.71 & 33.87 & 10.05 \\\\\nArchitecture $\\parallel$ 300-1024(-4096) & 17M & 64.29 & 58.05 & 48.49 & 33.94 & 10.09 \\\\\nArchitecture $\\parallel$ 300(-256-4096) & 4.5M & 62.82 & 56.28 & 48.02 & 32.71 & 9.89 \\\\\nArchitecture $\\parallel$ 300-4096 & 1.2M & 63.23 & 56.92 & 48.17 & 32.66 & 9.2 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Comparison of various bounding box regressors on Flickr30k Entities for different IoU thresholds. The number of parameters in Gr is also shown.\n
\n \n Params || -\n \n IoU || 0.5\n \n IoU || 0.6\n \n IoU || 0.7\n \n IoU || 0.8\n \n IoU || 0.9\n
\n Architecture || w/o regression\n \n -\n \n 65.21\n \n 53.19\n \n 35.70\n \n 14.32\n \n 1.88\n
\n Architecture || 300-16(-4096)\n \n 0.3M\n \n 64.14\n \n 57.66\n \n 48.22\n \n 33.04\n \n 9.29\n
\n Architecture || 300-64(-4096)\n \n 1.1M\n \n 63.87\n \n 57.43\n \n 49.05\n \n 33.84\n \n 10.55\n
\n Architecture || 300-256(-4096)\n \n 4.3M\n \n 63.84\n \n 57.70\n \n 48.71\n \n 33.87\n \n 10.05\n
\n Architecture || 300-1024(-4096)\n \n 17M\n \n 64.29\n \n 58.05\n \n 48.49\n \n 33.94\n \n 10.09\n
\n Architecture || 300(-256-4096)\n \n 4.5M\n \n 62.82\n \n 56.28\n \n 48.02\n \n 32.71\n \n 9.89\n
\n Architecture || 300-4096\n \n 1.2M\n \n 63.23\n \n 56.92\n \n 48.17\n \n 32.66\n \n 9.20\n
\n" }, { "instance_id": "7ded00a5-b726-4b4e-a235-b9e793070e1a", "table_id_paper": "D18-1303table_1", "caption": "Single-label classification (accuracy) results.", "row_header_level": 2, "row_headers": [ [ "Model", "Linear SVM" ], [ "Model", "Gaussian NB" ], [ "Model", "Logistic Reg." ], [ "Model", "SVM" ], [ "Model", "CNN" ], [ "Model", "RNN" ], [ "Model", "CNN-RNN" ] ], "column_header_level": 1, "column_headers": [ [ "Commenting" ], [ "Ogling" ], [ "Groping" ] ], "contents": [ [ "42.2", "35.0", "55.8" ], [ "46.8", "74.7", "66.0" ], [ "61.4", "78.0", "69.1" ], [ "65.5", "79.0", "70.3" ], [ "80.9", "82.2", "86.0" ], [ "81.0", "82.2", "86.2" ], [ "81.6", "84.1", "86.5" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy']", "target_entity": "['CNN-RNN']", "table_html": "
Commenting Ogling Groping
Model || Linear SVM 42.2 35.0 55.8
Model || Gaussian NB 46.8 74.7 66.0
Model || Logistic Reg. 61.4 78.0 69.1
Model || SVM 65.5 79.0 70.3
Model || CNN 80.9 82.2 86.0
Model || RNN 81.0 82.2 86.2
Model || CNN-RNN 81.6 84.1 86.5
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "D18-1303", "page_no": 3, "dir": "emnlp2018", "description": "5 Results. See Table 1 for single-label results on the selected harassment categories, where CNN-RNN was the best performing model compared to several nonneural and neural baselines.", "class_sentence": "[2, 1]", "sentences": "['5 Results.', 'See Table 1 for single-label results on the selected harassment categories, where CNN-RNN was the best performing model compared to several nonneural and neural baselines.']", "header_mention": "[None, ['CNN-RNN', 'Commenting', 'Ogling', 'Groping']]", "valid": 1, "image_id": "D18-1303-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Single-label classification (accuracy) results.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n CommentingOglingGroping
Model || Linear SVM42.235.055.8
Model || Gaussian NB46.874.766.0
Model || Logistic Reg.61.478.069.1
Model || SVM65.579.070.3
Model || CNN80.982.286.0
Model || RNN81.082.286.2
Model || CNN-RNN81.684.186.5
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Single-label classification (accuracy) results.}\n\\begin{tabular}{lrrr}\n\\toprule\n & Commenting & Ogling & Groping \\\\\n\\midrule\nModel $\\parallel$ Linear SVM & 42.2 & 35 & 55.8 \\\\\nModel $\\parallel$ Gaussian NB & 46.8 & 74.7 & 66 \\\\\nModel $\\parallel$ Logistic Reg. & 61.4 & 78 & 69.1 \\\\\nModel $\\parallel$ SVM & 65.5 & 79 & 70.3 \\\\\nModel $\\parallel$ CNN & 80.9 & 82.2 & 86 \\\\\nModel $\\parallel$ RNN & 81 & 82.2 & 86.2 \\\\\nModel $\\parallel$ CNN-RNN & 81.6 & 84.1 & 86.5 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Single-label classification (accuracy) results.\n
\n \n Commenting\n \n Ogling\n \n Groping\n
\n Model || Linear SVM\n \n 42.2\n \n 35.0\n \n 55.8\n
\n Model || Gaussian NB\n \n 46.8\n \n 74.7\n \n 66.0\n
\n Model || Logistic Reg.\n \n 61.4\n \n 78.0\n \n 69.1\n
\n Model || SVM\n \n 65.5\n \n 79.0\n \n 70.3\n
\n Model || CNN\n \n 80.9\n \n 82.2\n \n 86.0\n
\n Model || RNN\n \n 81.0\n \n 82.2\n \n 86.2\n
\n Model || CNN-RNN\n \n 81.6\n \n 84.1\n \n 86.5\n
\n" }, { "instance_id": "2b863c2e-8c94-4be7-8552-bdb4eb31e8db", "table_id_paper": "D18-1349table_6", "caption": "Results (presented in percentage) in terms of precision (P), recall (R) and F-measure (F1) on the test set for each label obtained by our HSLN-RNN model on the PubMed 20k dataset.", "row_header_level": 2, "row_headers": [ [ "Label", "Background" ], [ "Label", "Objectives" ], [ "Label", "Methods" ], [ "Label", "Results" ], [ "Label", "Conclusions" ], [ "Label", "Total" ] ], "column_header_level": 1, "column_headers": [ [ "P" ], [ "R" ], [ "F1" ], [ "Support" ] ], "contents": [ [ "78.5", "80.0", "79.2", "3077" ], [ "74.2", "69.9", "72.0", "2333" ], [ "95.0", "97.7", "96.3", "9884" ], [ "96.8", "95.3", "96.0", "9713" ], [ "97.6", "96.5", "97.1", "4571" ], [ "92.6", "92.7", "92.6", "29578" ] ], "metrics_loc": "column", "metrics_type": "['P', 'R', 'F1', 'Support']", "target_entity": "['Background', 'Objectives', 'Methods', 'Results', 'Conclusions']", "table_html": "
P R F1 Support
Label || Background 78.5 80.0 79.2 3077
Label || Objectives 74.2 69.9 72.0 2333
Label || Methods 95.0 97.7 96.3 9884
Label || Results 96.8 95.3 96.0 9713
Label || Conclusions 97.6 96.5 97.1 4571
Label || Total 92.6 92.7 92.6 29578
", "table_name": "Table 6", "table_id": "table_6", "paper_id": "D18-1349", "page_no": 7, "dir": "emnlp2018", "description": "Table 6 and 7 detail the results of classification for each label in terms of performance scores (precision, recall and F1) and confusion matrix, respectively (for our HSLN-RNN model trained on the PubMed 20k dataset). These show that the classifier is very good at predicting the labels Methods, Results and Conclusions, whereas the greatest difficulty the classifier has is in distinguishing Background sections from Objectives sections. One fifth of Background sentences are incorrectly classified as Objectives, while around one forth of Objectives sentences are wrongly assigned to the label of Background. We conjecture this difficulty mainly comes from the fact that the difference between Background and Objectives sentences in terms of writing style is less obvious compared with the other sections of the abstract. Moreover, our model has some difficulty in telling Methods sentences apart from Results sentences.", "class_sentence": "[1, 1, 2, 2, 2]", "sentences": "['Table 6 and 7 detail the results of classification for each label in terms of performance scores (precision, recall and F1) and confusion matrix, respectively (for our HSLN-RNN model trained on the PubMed 20k dataset).', 'These show that the classifier is very good at predicting the labels Methods, Results and Conclusions, whereas the greatest difficulty the classifier has is in distinguishing Background sections from Objectives sections.', 'One fifth of Background sentences are incorrectly classified as Objectives, while around one forth of Objectives sentences are wrongly assigned to the label of Background.', 'We conjecture this difficulty mainly comes from the fact that the difference between Background and Objectives sentences in terms of writing style is less obvious compared with the other sections of the abstract.', 'Moreover, our model has some difficulty in telling Methods sentences apart from Results sentences.']", "header_mention": "[['Background', 'Objectives', 'Methods', 'Results', 'Conclusions', 'P', 'R', 'F1'], ['Methods', 'Results', 'Conclusions', 'Background', 'Objectives'], ['Background', 'Objectives'], ['Background', 'Objectives'], ['Methods', 'Results']]", "valid": 1, "image_id": "D18-1349-Table6-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results (presented in percentage) in terms of precision (P), recall (R) and F-measure (F1) on the test set for each label obtained by our HSLN-RNN model on the PubMed 20k dataset.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n PRF1Support
Label || Background78.580.079.23077
Label || Objectives74.269.972.02333
Label || Methods95.097.796.39884
Label || Results96.895.396.09713
Label || Conclusions97.696.597.14571
Label || Total92.692.792.629578
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{5}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results (presented in percentage) in terms of precision (P), recall (R) and F-measure (F1) on the test set for each label obtained by our HSLN-RNN model on the PubMed 20k dataset.}\n\\begin{tabular}{lrrrr}\n\\toprule\n & P & R & F1 & Support \\\\\n\\midrule\nLabel $\\parallel$ Background & 78.5 & 80 & 79.2 & 3077 \\\\\nLabel $\\parallel$ Objectives & 74.2 & 69.9 & 72 & 2333 \\\\\nLabel $\\parallel$ Methods & 95 & 97.7 & 96.3 & 9884 \\\\\nLabel $\\parallel$ Results & 96.8 & 95.3 & 96 & 9713 \\\\\nLabel $\\parallel$ Conclusions & 97.6 & 96.5 & 97.1 & 4571 \\\\\nLabel $\\parallel$ Total & 92.6 & 92.7 & 92.6 & 29578 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 6: Results (presented in percentage) in terms of precision (P), recall (R) and F-measure (F1) on the test set for each label obtained by our HSLN-RNN model on the PubMed 20k dataset.\n
\n \n P\n \n R\n \n F1\n \n Support\n
\n Label || Background\n \n 78.5\n \n 80.0\n \n 79.2\n \n 3077\n
\n Label || Objectives\n \n 74.2\n \n 69.9\n \n 72.0\n \n 2333\n
\n Label || Methods\n \n 95.0\n \n 97.7\n \n 96.3\n \n 9884\n
\n Label || Results\n \n 96.8\n \n 95.3\n \n 96.0\n \n 9713\n
\n Label || Conclusions\n \n 97.6\n \n 96.5\n \n 97.1\n \n 4571\n
\n Label || Total\n \n 92.6\n \n 92.7\n \n 92.6\n \n 29578\n
\n" }, { "instance_id": "f423efc5-c56b-49cd-83de-d978a62407b0", "table_id_paper": "D18-1351table_4", "caption": "CV coherence scores for topics generated by various models. Higher is better. The best result in each column is in bold.", "row_header_level": 2, "row_headers": [ [ "Model", "LDA" ], [ "Model", "BTM" ], [ "Model", "NTM" ], [ "Model", "TMN" ] ], "column_header_level": 1, "column_headers": [ [ "Snippets" ], [ "TagMyNews" ], [ "Twitter" ] ], "contents": [ [ "0.436", "0.449", "0.436" ], [ "0.435", "0.463", "0.435" ], [ "0.463", "0.468", "0.463" ], [ "0.487", "0.499", "0.468" ] ], "metrics_loc": "column", "metrics_type": "['CV', 'CV', 'CV']", "target_entity": "['TMN']", "table_html": "
Snippets TagMyNews Twitter
Model || LDA 0.436 0.449 0.436
Model || BTM 0.435 0.463 0.435
Model || NTM 0.463 0.468 0.463
Model || TMN 0.487 0.499 0.468
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D18-1351", "page_no": 6, "dir": "emnlp2018", "description": "4.2 Topic Coherence Comparison In Section 4.1, we find that TMN can significantly outperform comparison models on short text classification. In this section, we study whether jointly learning topic models and classification can be helpful in producing coherent and meaningful topics. We use the CV metric (R\u00c2\u00a8oder et al., 2015) computed by Palmetto toolkit12 to evaluate the topic coherence, which has been shown to give the closest scores to human evaluation compared to other widely-used topic coherence metrics like NPMI (Bouma, 2009). Table 4 shows the comparison results of LDA, BTM, NTM, and TMN on the three English datasets. Note that we do not report CV scores for Chinese Weibo dataset as the Palmetto toolkit cannot process Chinese topics. As can be seen, TMN yields higher CV scores by large margins than all others in comparison. This indicates that jointly exploring classification would be effective in producing coherent topics. The reason is that the supervision from classification labels can guide unsupervised topic models in discovering meaningful and interpretable topics. We also observe that NTM produces better results than LDA and BTM, which implies the effectiveness of inducing topic models by neural networks.", "class_sentence": "[2, 2, 2, 1, 2, 1, 2, 2, 1]", "sentences": "['4.2 Topic Coherence Comparison In Section 4.1, we find that TMN can significantly outperform comparison models on short text classification.', 'In this section, we study whether jointly learning topic models and classification can be helpful in producing coherent and meaningful topics.', 'We use the CV metric (R\u00c2\u00a8oder et al., 2015) computed by Palmetto toolkit12 to evaluate the topic coherence, which has been shown to give the closest scores to human evaluation compared to other widely-used topic coherence metrics like NPMI (Bouma, 2009).', 'Table 4 shows the comparison results of LDA, BTM, NTM, and TMN on the three English datasets.', 'Note that we do not report CV scores for Chinese Weibo dataset as the Palmetto toolkit cannot process Chinese topics.', 'As can be seen, TMN yields higher CV scores by large margins than all others in comparison.', 'This indicates that jointly exploring classification would be effective in producing coherent topics.', 'The reason is that the supervision from classification labels can guide unsupervised topic models in discovering meaningful and interpretable topics.', 'We also observe that NTM produces better results than LDA and BTM, which implies the effectiveness of inducing topic models by neural networks.']", "header_mention": "[None, None, None, ['LDA', 'BTM', 'NTM', 'TMN', 'Snippets', 'TagMyNews', 'Twitter'], None, ['TMN'], ['TMN'], None, ['NTM', 'LDA', 'BTM']]", "valid": 1, "image_id": "D18-1351-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

CV coherence scores for topics generated by various models. Higher is better. The best result in each column is in bold.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n SnippetsTagMyNewsTwitter
Model || LDA0.4360.4490.436
Model || BTM0.4350.4630.435
Model || NTM0.4630.4680.463
Model || TMN0.4870.4990.468
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{CV coherence scores for topics generated by various models. Higher is better. The best result in each column is in bold.}\n\\begin{tabular}{lrrr}\n\\toprule\n & Snippets & TagMyNews & Twitter \\\\\n\\midrule\nModel $\\parallel$ LDA & 0.436 & 0.449 & 0.436 \\\\\nModel $\\parallel$ BTM & 0.435 & 0.463 & 0.435 \\\\\nModel $\\parallel$ NTM & 0.463 & 0.468 & 0.463 \\\\\nModel $\\parallel$ TMN & 0.487 & 0.499 & 0.468 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: CV coherence scores for topics generated by various models. Higher is better. The best result in each column is in bold.\n
\n \n Snippets\n \n TagMyNews\n \n Twitter\n
\n Model || LDA\n \n 0.436\n \n 0.449\n \n 0.436\n
\n Model || BTM\n \n 0.435\n \n 0.463\n \n 0.435\n
\n Model || NTM\n \n 0.463\n \n 0.468\n \n 0.463\n
\n Model || TMN\n \n 0.487\n \n 0.499\n \n 0.468\n
\n" }, { "instance_id": "57af7e89-e0c5-49de-9950-65ac8aa800cc", "table_id_paper": "D18-1362table_3", "caption": "Comparison of dev set MRR of Ours(ConvE) and models without reward shaping and action dropout.", "row_header_level": 2, "row_headers": [ [ "Model", "Ours(ConvE)" ], [ "Model", "-RS" ], [ "Model", "-AD" ] ], "column_header_level": 1, "column_headers": [ [ "UMLS" ], [ "Kinship" ], [ "FB15k237" ], [ "WN18RR" ], [ "NELL995" ] ], "contents": [ [ "73.0", "75.0", "38.2", "43.8", "78.8" ], [ "67.7", "66.5", "35.1", "45.7", "78.4" ], [ "61.3", "65.4", "31.0", "39.1", "76.1" ] ], "metrics_loc": "column", "metrics_type": "['MRR', 'MRR', 'MRR', 'MRR', 'MRR']", "target_entity": "['Ours(ConvE)', '-RS', '-AD']", "table_html": "
UMLS Kinship FB15k237 WN18RR NELL995
Model || Ours(ConvE) 73.0 75.0 38.2 43.8 78.8
Model || -RS 67.7 66.5 35.1 45.7 78.4
Model || -AD 61.3 65.4 31.0 39.1 76.1
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D18-1362", "page_no": 7, "dir": "emnlp2018", "description": "5.2 Ablation Study. We perform an ablation study where we remove reward shaping (\u2212RS) and action dropout (\u2212AD) from Ours(ConvE) and compare their MRRs to the whole model on the dev sets. As shown in Table 3, on most datasets, removing each component results in a significant performance drop. The exception is WN18RR, where removing the ConvE reward shaping module improves the performance. Removing reward shaping on NELL995 does not change the results significantly. In general, removing action dropout has a greater impact, suggesting that thorough exploration of the path space is important across datasets.", "class_sentence": "[2, 2, 1, 1, 1, 2]", "sentences": "['5.2 Ablation Study.', 'We perform an ablation study where we remove reward shaping (\u2212RS) and action dropout (\u2212AD) from Ours(ConvE) and compare their MRRs to the whole model on the dev sets.', 'As shown in Table 3, on most datasets, removing each component results in a significant performance drop.', 'The exception is WN18RR, where removing the ConvE reward shaping module improves the performance.', 'Removing reward shaping on NELL995 does not change the results significantly.', 'In general, removing action dropout has a greater impact, suggesting that thorough exploration of the path space is important across datasets.']", "header_mention": "[None, ['Ours(ConvE)', '-RS', '-AD'], ['-RS', '-AD'], ['WN18RR', '-RS', 'Ours(ConvE)'], ['NELL995', '-RS', 'Ours(ConvE)'], ['-AD']]", "valid": 1, "image_id": "D18-1362-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison of dev set MRR of Ours(ConvE) and models without reward shaping and action dropout.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n UMLSKinshipFB15k237WN18RRNELL995
Model || Ours(ConvE)73.075.038.243.878.8
Model || -RS67.766.535.145.778.4
Model || -AD61.365.431.039.176.1
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison of dev set MRR of Ours(ConvE) and models without reward shaping and action dropout.}\n\\begin{tabular}{lrrrrr}\n\\toprule\n & UMLS & Kinship & FB15k237 & WN18RR & NELL995 \\\\\n\\midrule\nModel $\\parallel$ Ours(ConvE) & 73 & 75 & 38.2 & 43.8 & 78.8 \\\\\nModel $\\parallel$ -RS & 67.7 & 66.5 & 35.1 & 45.7 & 78.4 \\\\\nModel $\\parallel$ -AD & 61.3 & 65.4 & 31 & 39.1 & 76.1 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Comparison of dev set MRR of Ours(ConvE) and models without reward shaping and action dropout.\n
\n \n UMLS\n \n Kinship\n \n FB15k237\n \n WN18RR\n \n NELL995\n
\n Model || Ours(ConvE)\n \n 73.0\n \n 75.0\n \n 38.2\n \n 43.8\n \n 78.8\n
\n Model || -RS\n \n 67.7\n \n 66.5\n \n 35.1\n \n 45.7\n \n 78.4\n
\n Model || -AD\n \n 61.3\n \n 65.4\n \n 31.0\n \n 39.1\n \n 76.1\n
\n" }, { "instance_id": "6b4a76c2-5316-4a40-8bfe-441372c27e10", "table_id_paper": "D18-1363table_3", "caption": "MED accuracy on five randomly selected languages with 1, 2, and 4 sources and combined with paradigm transduction (\u201c+PT\u201d). Best results in bold.", "row_header_level": 1, "row_headers": [ [ "dutch" ], [ "german" ], [ "icelandic" ], [ "spanish" ], [ "welsh" ] ], "column_header_level": 2, "column_headers": [ [ "SET1", "1" ], [ "SET1", "2" ], [ "SET1", "4" ], [ "SET1", "+PT" ], [ "SET2", "1" ], [ "SET2", "2" ], [ "SET2", "4" ], [ "SET2", "+PT" ], [ "SET3", "1" ], [ "SET3", "2" ], [ "SET3", "4" ], [ "SET3", "+PT" ] ], "contents": [ [ ".00", ".00", ".00", ".49", ".04", ".01", ".00", ".78", ".43", ".65", ".72", ".87" ], [ ".00", ".00", ".00", ".65", ".00", ".00", ".01", ".75", ".44", ".42", ".59", ".88" ], [ ".00", ".00", ".00", ".41", ".03", ".02", ".02", ".50", ".24", ".33", ".35", ".77" ], [ ".00", ".00", ".00", ".92", ".03", ".09", ".09", ".98", ".59", ".63", ".83", ".99" ], [ ".00", ".00", ".00", ".91", ".05", ".14", ".15", ".97", ".35", ".53", ".70", ".99" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy']", "target_entity": "['1', '2', '4']", "table_html": "
SET1 || 1 SET1 || 2 SET1 || 4 SET1 || +PT SET2 || 1 SET2 || 2 SET2 || 4 SET2 || +PT SET3 || 1 SET3 || 2 SET3 || 4 SET3 || +PT
dutch .00 .00 .00 .49 .04 .01 .00 .78 .43 .65 .72 .87
german .00 .00 .00 .65 .00 .00 .01 .75 .44 .42 .59 .88
icelandic .00 .00 .00 .41 .03 .02 .02 .50 .24 .33 .35 .77
spanish .00 .00 .00 .92 .03 .09 .09 .98 .59 .63 .83 .99
welsh .00 .00 .00 .91 .05 .14 .15 .97 .35 .53 .70 .99
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D18-1363", "page_no": 8, "dir": "emnlp2018", "description": "The input to a multi-source model is the concatenation of all sources and corresponding tags. During training, we randomly sample (with repetition) one or three additional forms from the paradigm of each example. At test time, we sample the additional forms from the given partial paradigm; without repetition first, but repeating if not enough inflected forms are available. For autoencoding examples in the training data, we simply concatenate two or four copies of the source and the autoencoding tag. We randomly select five languages for this experiment. Table 3 shows that, for SET3, four sources (column header\"4\") are generally better than two sources (\"2\"), which in turn are better than one source (\"1\"); thus, as expected, making additional sources available in training improves results. We attribute one exception (German accuracy is .4391 for \"1\" and .4179 for \"2\") to the noisiness of the problem training sets in terms of number of paradigms are relatively small, even for SET3.", "class_sentence": "[2, 2, 2, 2, 2, 1, 2]", "sentences": "['The input to a multi-source model is the concatenation of all sources and corresponding tags.', 'During training, we randomly sample (with repetition) one or three additional forms from the paradigm of each example.', 'At test time, we sample the additional forms from the given partial paradigm; without repetition first, but repeating if not enough inflected forms are available.', 'For autoencoding examples in the training data, we simply concatenate two or four copies of the source and the autoencoding tag.', 'We randomly select five languages for this experiment.', 'Table 3 shows that, for SET3, four sources (column header\"4\") are generally better than two sources (\"2\"), which in turn are better than one source (\"1\"); thus, as expected, making additional sources available in training improves results.', 'We attribute one exception (German accuracy is .4391 for \"1\" and .4179 for \"2\") to the noisiness of the problem training sets in terms of number of paradigms are relatively small, even for SET3.']", "header_mention": "[None, None, None, None, None, ['SET3', '1', '2', '4'], ['german']]", "valid": 1, "image_id": "D18-1363-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

MED accuracy on five randomly selected languages with 1, 2, and 4 sources and combined with paradigm transduction (\u201c+PT\u201d). Best results in bold.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n SET1 || 1SET1 || 2SET1 || 4SET1 || +PTSET2 || 1SET2 || 2SET2 || 4SET2 || +PTSET3 || 1SET3 || 2SET3 || 4SET3 || +PT
dutch.00.00.00.49.04.01.00.78.43.65.72.87
german.00.00.00.65.00.00.01.75.44.42.59.88
icelandic.00.00.00.41.03.02.02.50.24.33.35.77
spanish.00.00.00.92.03.09.09.98.59.63.83.99
welsh.00.00.00.91.05.14.15.97.35.53.70.99
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{MED accuracy on five randomly selected languages with 1, 2, and 4 sources and combined with paradigm transduction (\u201c+PT\u201d). Best results in bold.}\n\\begin{tabular}{lrrrrrrrrrrrr}\n\\toprule\n & SET1 $\\parallel$ 1 & SET1 $\\parallel$ 2 & SET1 $\\parallel$ 4 & SET1 $\\parallel$ +PT & SET2 $\\parallel$ 1 & SET2 $\\parallel$ 2 & SET2 $\\parallel$ 4 & SET2 $\\parallel$ +PT & SET3 $\\parallel$ 1 & SET3 $\\parallel$ 2 & SET3 $\\parallel$ 4 & SET3 $\\parallel$ +PT \\\\\n\\midrule\ndutch & 0 & 0 & 0 & 0.49 & 0.04 & 0.01 & 0 & 0.78 & 0.43 & 0.65 & 0.72 & 0.87 \\\\\ngerman & 0 & 0 & 0 & 0.65 & 0 & 0 & 0.01 & 0.75 & 0.44 & 0.42 & 0.59 & 0.88 \\\\\nicelandic & 0 & 0 & 0 & 0.41 & 0.03 & 0.02 & 0.02 & 0.5 & 0.24 & 0.33 & 0.35 & 0.77 \\\\\nspanish & 0 & 0 & 0 & 0.92 & 0.03 & 0.09 & 0.09 & 0.98 & 0.59 & 0.63 & 0.83 & 0.99 \\\\\nwelsh & 0 & 0 & 0 & 0.91 & 0.05 & 0.14 & 0.15 & 0.97 & 0.35 & 0.53 & 0.7 & 0.99 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: MED accuracy on five randomly selected languages with 1, 2, and 4 sources and combined with paradigm transduction (\u201c+PT\u201d). Best results in bold.\n
\n \n SET1 || 1\n \n SET1 || 2\n \n SET1 || 4\n \n SET1 || +PT\n \n SET2 || 1\n \n SET2 || 2\n \n SET2 || 4\n \n SET2 || +PT\n \n SET3 || 1\n \n SET3 || 2\n \n SET3 || 4\n \n SET3 || +PT\n
\n dutch\n \n .00\n \n .00\n \n .00\n \n .49\n \n .04\n \n .01\n \n .00\n \n .78\n \n .43\n \n .65\n \n .72\n \n .87\n
\n german\n \n .00\n \n .00\n \n .00\n \n .65\n \n .00\n \n .00\n \n .01\n \n .75\n \n .44\n \n .42\n \n .59\n \n .88\n
\n icelandic\n \n .00\n \n .00\n \n .00\n \n .41\n \n .03\n \n .02\n \n .02\n \n .50\n \n .24\n \n .33\n \n .35\n \n .77\n
\n spanish\n \n .00\n \n .00\n \n .00\n \n .92\n \n .03\n \n .09\n \n .09\n \n .98\n \n .59\n \n .63\n \n .83\n \n .99\n
\n welsh\n \n .00\n \n .00\n \n .00\n \n .91\n \n .05\n \n .14\n \n .15\n \n .97\n \n .35\n \n .53\n \n .70\n \n .99\n
\n" }, { "instance_id": "c14a60f2-b7f3-4b32-9a2c-ee50fa4f9399", "table_id_paper": "D18-1368table_2", "caption": "Situation Entity Type Classification Results on the Training Set of MASC+Wiki with 10-Fold Cross-Validation. We report accuracy (Acc), macro-average F1-score (Macro) and class-wise F1 scores for STATE (STA), EVENT (EVE), REPORT (REP), GENERIC (GENI), GENERALIZING (GENA), QUESTION (QUE) and IMPERATIVE (IMP).", "row_header_level": 2, "row_headers": [ [ "Model", "Humans" ], [ "Model", "CRF (Friedrich et al. 2016)" ], [ "Model", "Clause-level Bi-LSTM" ], [ "Model", "Paragraph-level Model" ], [ "Model", "Paragraph-level Model+CRF" ] ], "column_header_level": 1, "column_headers": [ [ "Macro" ], [ "Acc" ], [ "F1 STA" ], [ "F1 EVE" ], [ "F1 REP" ], [ "F1 GENI" ], [ "F1 GENA" ], [ "F1 QUE" ], [ "F1 IMP" ] ], "contents": [ [ "78.6", "79.6", "82.8", "80.5", "81.5", "75.1", "45.8", "90.7", "93.6" ], [ "71.2", "76.4", "80.6", "78.6", "78.9", "68.3", "29.4", "84.4", "75.3" ], [ "74.4", "78.3", "82.6", "81.3", "84.9", "66.2", "36.1", "88.5", "80.9" ], [ "77.6", "81.2", "84.3", "82.1", "85.3", "76.4", "43.2", "90.8", "81.2" ], [ "77.8", "81.3", "84.3", "82.0", "85.7", "77.0", "43.5", "90.4", "81.5" ] ], "metrics_loc": "column", "metrics_type": "['Macro', 'Acc', 'F1 STA', 'F1 EVE', 'F1 REP', 'F1 GENI', 'F1 GENA', 'F1 QUE', 'F1 IMP']", "target_entity": "['Paragraph-level Model', 'Paragraph-level Model+CRF']", "table_html": "
Macro Acc STA EVE REP GENI GENA QUE IMP
Model || Humans 78.6 79.6 82.8 80.5 81.5 75.1 45.8 90.7 93.6
Model || CRF (Friedrich et al. 2016) 71.2 76.4 80.6 78.6 78.9 68.3 29.4 84.4 75.3
Model || Clause-level Bi-LSTM 74.4 78.3 82.6 81.3 84.9 66.2 36.1 88.5 80.9
Model || Paragraph-level Model 77.6 81.2 84.3 82.1 85.3 76.4 43.2 90.8 81.2
Model || Paragraph-level Model+CRF 77.8 81.3 84.3 82.0 85.7 77.0 43.5 90.4 81.5
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1368", "page_no": 6, "dir": "emnlp2018", "description": "Table 4 shows cross-genre experimental results of our neural network models on the training set of MASC+Wiki by treating each genre as one crossvalidation fold. As we expected, both the macroaverage F1-score and class-wise F1 scores are lower compared with the results in Table 2 where in-genre data were used for model training as well. But the performance drop on the paragraph-level models is little, which clearly outperform the previous system (Friedrich et al., 2016) and the baseline model by a large margin. As shown in Table 5, benefited from modeling wider contexts and common SE label patterns, our full paragraphlevel model improves performance across almost all the genres. The high performance in the crossgenre setting demonstrates the robustness of our paragraph-level model across genres. Table 2 reports the cross-validation classification results. Consistently, our clause-level baseline model already outperforms the previous best model. By exploiting paragraph-wide contexts, the basic paragraph-level model obtains consistent performance improvements across all the classes compared with the baseline clause-level prediction model, especially for the classes GENERIC and GENERALIZING, where the improvements are significant. After using the CRF layer to fine-tune the predicted SE label sequence, slight performance improvements were observed on the four small classes. Overall, the full paragraphlevel neural network model achieves the best macro-average F1-score of 77.8% in predicting SE types, which not only outperforms all previous approaches but also reaches human-like performance on some classes.", "class_sentence": "[0, 2, 0, 0, 0, 1, 1, 1, 1, 1]", "sentences": "['Table 4 shows cross-genre experimental results of our neural network models on the training set of MASC+Wiki by treating each genre as one crossvalidation fold.', 'As we expected, both the macroaverage F1-score and class-wise F1 scores are lower compared with the results in Table 2 where in-genre data were used for model training as well.', 'But the performance drop on the paragraph-level models is little, which clearly outperform the previous system (Friedrich et al., 2016) and the baseline model by a large margin.', 'As shown in Table 5, benefited from modeling wider contexts and common SE label patterns, our full paragraphlevel model improves performance across almost all the genres.', 'The high performance in the crossgenre setting demonstrates the robustness of our paragraph-level model across genres.', 'Table 2 reports the cross-validation classification results.', 'Consistently, our clause-level baseline model already outperforms the previous best model.', 'By exploiting paragraph-wide contexts, the basic paragraph-level model obtains consistent performance improvements across all the classes compared with the baseline clause-level prediction model, especially for the classes GENERIC and GENERALIZING, where the improvements are significant.', 'After using the CRF layer to fine-tune the predicted SE label sequence, slight performance improvements were observed on the four small classes.', 'Overall, the full paragraphlevel neural network model achieves the best macro-average F1-score of 77.8% in predicting SE types, which not only outperforms all previous approaches but also reaches human-like performance on some classes.']", "header_mention": "[None, None, None, None, None, None, ['Clause-level Bi-LSTM', 'CRF (Friedrich et al. 2016)'], ['Clause-level Bi-LSTM', 'Paragraph-level Model', 'F1 GENI', 'F1 GENA'], ['Paragraph-level Model', 'Paragraph-level Model+CRF', 'F1 REP', 'F1 GENI', 'F1 GENA', 'F1 IMP'], ['Paragraph-level Model+CRF', 'Macro', 'Humans']]", "valid": 1, "image_id": "D18-1368-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Situation Entity Type Classification Results on the Training Set of MASC+Wiki with 10-Fold Cross-Validation. We report accuracy (Acc), macro-average F1-score (Macro) and class-wise F1 scores for STATE (STA), EVENT (EVE), REPORT (REP), GENERIC (GENI), GENERALIZING (GENA), QUESTION (QUE) and IMPERATIVE (IMP).

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n MacroAccSTAEVEREPGENIGENAQUEIMP
Model || Humans78.679.682.880.581.575.145.890.793.6
Model || CRF (Friedrich et al. 2016)71.276.480.678.678.968.329.484.475.3
Model || Clause-level Bi-LSTM74.478.382.681.384.966.236.188.580.9
Model || Paragraph-level Model77.681.284.382.185.376.443.290.881.2
Model || Paragraph-level Model+CRF77.881.384.382.085.777.043.590.481.5
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Situation Entity Type Classification Results on the Training Set of MASC+Wiki with 10-Fold Cross-Validation. We report accuracy (Acc), macro-average F1-score (Macro) and class-wise F1 scores for STATE (STA), EVENT (EVE), REPORT (REP), GENERIC (GENI), GENERALIZING (GENA), QUESTION (QUE) and IMPERATIVE (IMP).}\n\\begin{tabular}{lrrrrrrrrr}\n\\toprule\n & Macro & Acc & STA & EVE & REP & GENI & GENA & QUE & IMP \\\\\n\\midrule\nModel $\\parallel$ Humans & 78.6 & 79.6 & 82.8 & 80.5 & 81.5 & 75.1 & 45.8 & 90.7 & 93.6 \\\\\nModel $\\parallel$ CRF (Friedrich et al. 2016) & 71.2 & 76.4 & 80.6 & 78.6 & 78.9 & 68.3 & 29.4 & 84.4 & 75.3 \\\\\nModel $\\parallel$ Clause-level Bi-LSTM & 74.4 & 78.3 & 82.6 & 81.3 & 84.9 & 66.2 & 36.1 & 88.5 & 80.9 \\\\\nModel $\\parallel$ Paragraph-level Model & 77.6 & 81.2 & 84.3 & 82.1 & 85.3 & 76.4 & 43.2 & 90.8 & 81.2 \\\\\nModel $\\parallel$ Paragraph-level Model+CRF & 77.8 & 81.3 & 84.3 & 82 & 85.7 & 77 & 43.5 & 90.4 & 81.5 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Situation Entity Type Classification Results on the Training Set of MASC+Wiki with 10-Fold Cross-Validation. We report accuracy (Acc), macro-average F1-score (Macro) and class-wise F1 scores for STATE (STA), EVENT (EVE), REPORT (REP), GENERIC (GENI), GENERALIZING (GENA), QUESTION (QUE) and IMPERATIVE (IMP).\n
\n \n Macro\n \n Acc\n \n STA\n \n EVE\n \n REP\n \n GENI\n \n GENA\n \n QUE\n \n IMP\n
\n Model || Humans\n \n 78.6\n \n 79.6\n \n 82.8\n \n 80.5\n \n 81.5\n \n 75.1\n \n 45.8\n \n 90.7\n \n 93.6\n
\n Model || CRF (Friedrich et al. 2016)\n \n 71.2\n \n 76.4\n \n 80.6\n \n 78.6\n \n 78.9\n \n 68.3\n \n 29.4\n \n 84.4\n \n 75.3\n
\n Model || Clause-level Bi-LSTM\n \n 74.4\n \n 78.3\n \n 82.6\n \n 81.3\n \n 84.9\n \n 66.2\n \n 36.1\n \n 88.5\n \n 80.9\n
\n Model || Paragraph-level Model\n \n 77.6\n \n 81.2\n \n 84.3\n \n 82.1\n \n 85.3\n \n 76.4\n \n 43.2\n \n 90.8\n \n 81.2\n
\n Model || Paragraph-level Model+CRF\n \n 77.8\n \n 81.3\n \n 84.3\n \n 82.0\n \n 85.7\n \n 77.0\n \n 43.5\n \n 90.4\n \n 81.5\n
\n" }, { "instance_id": "795b77df-af59-4e56-9a57-b7caffd97ec0", "table_id_paper": "D18-1369table_2", "caption": "Narrative reconstruction results in NYT dataset and post grouping results in Wikipedia conversation dataset.", "row_header_level": 2, "row_headers": [ [ "NYT", "LDA + DBSCAN" ], [ "NYT", "HDP + DBSCAN" ], [ "NYT", "HDHP" ], [ "NYT", "HD-GMHP (100D)" ], [ "Wiki", "W2V + DBSCAN" ], [ "Wiki", "HDHP" ], [ "Wiki", "HD-GMHP (100D)" ] ], "column_header_level": 1, "column_headers": [ [ "AMI" ], [ "ARI" ] ], "contents": [ [ "0.0627", "0.0117" ], [ "0.0260", "0.0203" ], [ "0.1768", "0.0746" ], [ "0.2479", "0.1416" ], [ "0.0055", "0.0001" ], [ "0.4240", "0.3512" ], [ "0.5848", "0.3834" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy']", "target_entity": "['HD-GMHP (100D)']", "table_html": "
AMI ARI
NYT || LDA + DBSCAN 0.0627 0.0117
NYT || HDP + DBSCAN 0.0260 0.0203
NYT || HDHP 0.1768 0.0746
NYT || HD-GMHP (100D) 0.2479 0.1416
Wiki || W2V + DBSCAN 0.0055 0.0001
Wiki || HDHP 0.4240 0.3512
Wiki || HD-GMHP (100D) 0.5848 0.3834
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1369", "page_no": 8, "dir": "emnlp2018", "description": "7.5 Results. Table 2 shows the clustering accuracy of our method and the baseline methods in real world datasets. We average the results with five runs for each model. The highest value for each metric is indicated with boldface. From the results, we establish that our model outperforms the baseline methods in both the NYT narrative reconstruction task and the Wikipedia thread reconstruction task. For the NYT, to see the accuracy of our model in more detail, we compute and show the F-scores for the top ten most frequent labels and the micro and macro averages in table 3. To compute the F-score between the true labels and the recovered cluster labels, we select the cluster with the highest F-score as the corresponding cluster. From the results, we establish that our model performs better than the baseline model, HDHP.", "class_sentence": "[2, 1, 2, 2, 1, 2, 2, 1]", "sentences": "['7.5 Results.', 'Table 2 shows the clustering accuracy of our method and the baseline methods in real world datasets.', 'We average the results with five runs for each model.', 'The highest value for each metric is indicated with boldface.', 'From the results, we establish that our model outperforms the baseline methods in both the NYT narrative reconstruction task and the Wikipedia thread reconstruction task.', 'For the NYT, to see the accuracy of our model in more detail, we compute and show the F-scores for the top ten most frequent labels and the micro and macro averages in table 3.', 'To compute the F-score between the true labels and the recovered cluster labels, we select the cluster with the highest F-score as the corresponding cluster.', 'From the results, we establish that our model performs better than the baseline model, HDHP.']", "header_mention": "[None, None, None, None, ['HD-GMHP (100D)', 'NYT', 'Wiki'], ['HD-GMHP (100D)', 'NYT'], None, ['HD-GMHP (100D)']]", "valid": 1, "image_id": "D18-1369-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Narrative reconstruction results in NYT dataset and post grouping results in Wikipedia conversation dataset.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n AMIARI
NYT || LDA + DBSCAN0.06270.0117
NYT || HDP + DBSCAN0.02600.0203
NYT || HDHP0.17680.0746
NYT || HD-GMHP (100D)0.24790.1416
Wiki || W2V + DBSCAN0.00550.0001
Wiki || HDHP0.42400.3512
Wiki || HD-GMHP (100D)0.58480.3834
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Narrative reconstruction results in NYT dataset and post grouping results in Wikipedia conversation dataset.}\n\\begin{tabular}{lrr}\n\\toprule\n & AMI & ARI \\\\\n\\midrule\nNYT $\\parallel$ LDA + DBSCAN & 0.0627 & 0.0117 \\\\\nNYT $\\parallel$ HDP + DBSCAN & 0.026 & 0.0203 \\\\\nNYT $\\parallel$ HDHP & 0.1768 & 0.0746 \\\\\nNYT $\\parallel$ HD-GMHP (100D) & 0.2479 & 0.1416 \\\\\nWiki $\\parallel$ W2V + DBSCAN & 0.0055 & 0.0001 \\\\\nWiki $\\parallel$ HDHP & 0.424 & 0.3512 \\\\\nWiki $\\parallel$ HD-GMHP (100D) & 0.5848 & 0.3834 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Narrative reconstruction results in NYT dataset and post grouping results in Wikipedia conversation dataset.\n
\n \n AMI\n \n ARI\n
\n NYT || LDA + DBSCAN\n \n 0.0627\n \n 0.0117\n
\n NYT || HDP + DBSCAN\n \n 0.0260\n \n 0.0203\n
\n NYT || HDHP\n \n 0.1768\n \n 0.0746\n
\n NYT || HD-GMHP (100D)\n \n 0.2479\n \n 0.1416\n
\n Wiki || W2V + DBSCAN\n \n 0.0055\n \n 0.0001\n
\n Wiki || HDHP\n \n 0.4240\n \n 0.3512\n
\n Wiki || HD-GMHP (100D)\n \n 0.5848\n \n 0.3834\n
\n" }, { "instance_id": "ea6a25ca-cfba-4d6b-8cec-1fca55de56d9", "table_id_paper": "D18-1381table_2", "caption": "Experimental results on Sem2016 with two training settings TRAIN and TRAIN-ALL.", "row_header_level": 1, "row_headers": [ [ "NBOW-MLP" ], [ "CNN" ], [ "BiLSTM" ], [ "AT-BiLSTM" ], [ "Lexicon RNN" ], [ "AGLR" ] ], "column_header_level": 3, "column_headers": [ [ "Sem2016 (TRAIN)", "3-way", "Acc" ], [ "Sem2016 (TRAIN)", "3-way", "F1" ], [ "Sem2016 (TRAIN)", "Binary", "Acc" ], [ "Sem2016 (TRAIN)", "Binary", "F1" ], [ "Sem2016 (TRAIN-ALL)", "3-way", "Acc" ], [ "Sem2016 (TRAIN-ALL)", "3-way", "F1" ], [ "Sem2016 (TRAIN-ALL)", "Binary", "Acc" ], [ "Sem2016 (TRAIN-ALL)", "Binary", "F1" ], [ "-", "AVG", "Acc" ], [ "-", "AVG", "F1" ] ], "contents": [ [ "54.31", "52.90", "79.69", "77.33", "61.09", "55.91", "84.90", "82.01", "70.00", "67.04" ], [ "54.67", "52.17", "79.79", "75.28", "62.68", "57.71", "84.10", "81.31", "70.31", "66.62" ], [ "55.57", "52.33", "81.90", "77.12", "63.26", "60.31", "85.89", "84.14", "71.66", "68.50" ], [ "56.95", "54.53", "80.09", "73.93", "64.20", "61.64", "86.77", "83.67", "72.00", "68.44" ], [ "51.02", "50.45", "81.72", "79.00", "61.41", "60.50", "86.68", "83.82", "70.21", "68.44" ], [ "59.01", "56.67", "82.79", "80.10", "66.62", "64.36", "87.16", "84.98", "73.90", "71.53" ] ], "metrics_loc": "column", "metrics_type": "['Acc', 'F1', 'Acc', 'F1', 'Acc', 'F1', 'Acc', 'F1', 'Acc', 'F1']", "target_entity": "['AGLR']", "table_html": "
Sem2016 (TRAIN) || 3-way || Acc Sem2016 (TRAIN) || 3-way || F1 Sem2016 (TRAIN) || Binary || Acc Sem2016 (TRAIN) || Binary || F1 Sem2016 (TRAIN-ALL) || 3-way || Acc Sem2016 (TRAIN-ALL) || 3-way || F1 Sem2016 (TRAIN-ALL) || Binary || Acc Sem2016 (TRAIN-ALL) || Binary || F1 - || AVG || Acc - || AVG || F1
NBOW-MLP 54.31 52.90 79.69 77.33 61.09 55.91 84.90 82.01 70.00 67.04
CNN 54.67 52.17 79.79 75.28 62.68 57.71 84.10 81.31 70.31 66.62
BiLSTM 55.57 52.33 81.90 77.12 63.26 60.31 85.89 84.14 71.66 68.50
AT-BiLSTM 56.95 54.53 80.09 73.93 64.20 61.64 86.77 83.67 72.00 68.44
Lexicon RNN 51.02 50.45 81.72 79.00 61.41 60.50 86.68 83.82 70.21 68.44
AGLR 59.01 56.67 82.79 80.10 66.62 64.36 87.16 84.98 73.90 71.53
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1381", "page_no": 7, "dir": "emnlp2018", "description": "4.2 Experimental Results. Table 1 and Table 2 report the results of our experiments. The results on TRAIN-ALL are higher than TRAIN for SemEval16 in lieu of the larger dataset. Firstly, we observe that our proposed AGLR outperforms all neural baselines on 3-way classification. The overall performance of AGLR achieves state-of-the-art performance. On average, AGLR outperforms Lexicon RNN and AT-BiLSTM by 1% \u00e2\u02c6\u2019 3% in terms of F1 score. We also observe that AGLR always improves AT-BiLSTM which ascertains the effectiveness of learning auxiliary lexicon embeddings. The key idea here is that the auxiliary lexicon embeddings provide a different view of the sentence which supports the network in making predictions. We also observe that Lexicon RNN does not handle 3-way classification well. Even though it has achieved good performance on binary classification, the performance on 3-way classification is lackluster (the performance of AGLR outperforms Lexicon RNN by up to 8% on SemEval16 TRAIN). This could also be attributed to the MSE based loss function. Conversely, by learning an auxiliary embedding (instead of a scalar score), our model becomes more flexible at the final layer and can be adapted to using a k softmax function. Finally, we observe that BiLSTM and AT-BiLSTM outperform Lexicon RNN on average with Lexicon RNN being slightly better on binary classification.", "class_sentence": "[2, 1, 1, 1, 1, 1, 1, 2, 1, 1, 2, 2, 1]", "sentences": "['4.2 Experimental Results.', 'Table 1 and Table 2 report the results of our experiments.', 'The results on TRAIN-ALL are higher than TRAIN for SemEval16 in lieu of the larger dataset.', 'Firstly, we observe that our proposed AGLR outperforms all neural baselines on 3-way classification.', 'The overall performance of AGLR achieves state-of-the-art performance.', 'On average, AGLR outperforms Lexicon RNN and AT-BiLSTM by 1% \u00e2\u02c6\u2019 3% in terms of F1 score.', 'We also observe that AGLR always improves AT-BiLSTM which ascertains the effectiveness of learning auxiliary lexicon embeddings.', 'The key idea here is that the auxiliary lexicon embeddings provide a different view of the sentence which supports the network in making predictions.', 'We also observe that Lexicon RNN does not handle 3-way classification well.', 'Even though it has achieved good performance on binary classification, the performance on 3-way classification is lackluster (the performance of AGLR outperforms Lexicon RNN by up to 8% on SemEval16 TRAIN).', 'This could also be attributed to the MSE based loss function.', 'Conversely, by learning an auxiliary embedding (instead of a scalar score), our model becomes more flexible at the final layer and can be adapted to using a k softmax function.', 'Finally, we observe that BiLSTM and AT-BiLSTM outperform Lexicon RNN on average with Lexicon RNN being slightly better on binary classification.']", "header_mention": "[None, None, ['Sem2016 (TRAIN)', 'Sem2016 (TRAIN-ALL)'], ['AGLR', '3-way'], ['AGLR'], ['AGLR', 'AT-BiLSTM', 'Lexicon RNN', 'AVG', 'F1'], ['AGLR', 'AT-BiLSTM'], None, ['Lexicon RNN', '3-way'], ['Lexicon RNN', 'Binary', 'AGLR', 'Sem2016 (TRAIN)'], None, ['AGLR'], ['BiLSTM', 'AT-BiLSTM', 'Lexicon RNN', 'Binary', 'AVG']]", "valid": 1, "image_id": "D18-1381-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Experimental results on Sem2016 with two training settings TRAIN and TRAIN-ALL.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Sem2016 (TRAIN) || 3-way || AccSem2016 (TRAIN) || 3-way || F1Sem2016 (TRAIN) || Binary || AccSem2016 (TRAIN) || Binary || F1Sem2016 (TRAIN-ALL) || 3-way || AccSem2016 (TRAIN-ALL) || 3-way || F1Sem2016 (TRAIN-ALL) || Binary || AccSem2016 (TRAIN-ALL) || Binary || F1- || AVG || Acc- || AVG || F1
NBOW-MLP54.3152.9079.6977.3361.0955.9184.9082.0170.0067.04
CNN54.6752.1779.7975.2862.6857.7184.1081.3170.3166.62
BiLSTM55.5752.3381.9077.1263.2660.3185.8984.1471.6668.50
AT-BiLSTM56.9554.5380.0973.9364.2061.6486.7783.6772.0068.44
Lexicon RNN51.0250.4581.7279.0061.4160.5086.6883.8270.2168.44
AGLR59.0156.6782.7980.1066.6264.3687.1684.9873.9071.53
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Experimental results on Sem2016 with two training settings TRAIN and TRAIN-ALL.}\n\\begin{tabular}{lrrrrrrrrrr}\n\\toprule\n & Sem2016 (TRAIN) $\\parallel$ 3-way $\\parallel$ Acc & Sem2016 (TRAIN) $\\parallel$ 3-way $\\parallel$ F1 & Sem2016 (TRAIN) $\\parallel$ Binary $\\parallel$ Acc & Sem2016 (TRAIN) $\\parallel$ Binary $\\parallel$ F1 & Sem2016 (TRAIN-ALL) $\\parallel$ 3-way $\\parallel$ Acc & Sem2016 (TRAIN-ALL) $\\parallel$ 3-way $\\parallel$ F1 & Sem2016 (TRAIN-ALL) $\\parallel$ Binary $\\parallel$ Acc & Sem2016 (TRAIN-ALL) $\\parallel$ Binary $\\parallel$ F1 & - $\\parallel$ AVG $\\parallel$ Acc & - $\\parallel$ AVG $\\parallel$ F1 \\\\\n\\midrule\nNBOW-MLP & 54.31 & 52.9 & 79.69 & 77.33 & 61.09 & 55.91 & 84.9 & 82.01 & 70 & 67.04 \\\\\nCNN & 54.67 & 52.17 & 79.79 & 75.28 & 62.68 & 57.71 & 84.1 & 81.31 & 70.31 & 66.62 \\\\\nBiLSTM & 55.57 & 52.33 & 81.9 & 77.12 & 63.26 & 60.31 & 85.89 & 84.14 & 71.66 & 68.5 \\\\\nAT-BiLSTM & 56.95 & 54.53 & 80.09 & 73.93 & 64.2 & 61.64 & 86.77 & 83.67 & 72 & 68.44 \\\\\nLexicon RNN & 51.02 & 50.45 & 81.72 & 79 & 61.41 & 60.5 & 86.68 & 83.82 & 70.21 & 68.44 \\\\\nAGLR & 59.01 & 56.67 & 82.79 & 80.1 & 66.62 & 64.36 & 87.16 & 84.98 & 73.9 & 71.53 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Experimental results on Sem2016 with two training settings TRAIN and TRAIN-ALL.\n
\n \n Sem2016 (TRAIN) || 3-way || Acc\n \n Sem2016 (TRAIN) || 3-way || F1\n \n Sem2016 (TRAIN) || Binary || Acc\n \n Sem2016 (TRAIN) || Binary || F1\n \n Sem2016 (TRAIN-ALL) || 3-way || Acc\n \n Sem2016 (TRAIN-ALL) || 3-way || F1\n \n Sem2016 (TRAIN-ALL) || Binary || Acc\n \n Sem2016 (TRAIN-ALL) || Binary || F1\n \n - || AVG || Acc\n \n - || AVG || F1\n
\n NBOW-MLP\n \n 54.31\n \n 52.90\n \n 79.69\n \n 77.33\n \n 61.09\n \n 55.91\n \n 84.90\n \n 82.01\n \n 70.00\n \n 67.04\n
\n CNN\n \n 54.67\n \n 52.17\n \n 79.79\n \n 75.28\n \n 62.68\n \n 57.71\n \n 84.10\n \n 81.31\n \n 70.31\n \n 66.62\n
\n BiLSTM\n \n 55.57\n \n 52.33\n \n 81.90\n \n 77.12\n \n 63.26\n \n 60.31\n \n 85.89\n \n 84.14\n \n 71.66\n \n 68.50\n
\n AT-BiLSTM\n \n 56.95\n \n 54.53\n \n 80.09\n \n 73.93\n \n 64.20\n \n 61.64\n \n 86.77\n \n 83.67\n \n 72.00\n \n 68.44\n
\n Lexicon RNN\n \n 51.02\n \n 50.45\n \n 81.72\n \n 79.00\n \n 61.41\n \n 60.50\n \n 86.68\n \n 83.82\n \n 70.21\n \n 68.44\n
\n AGLR\n \n 59.01\n \n 56.67\n \n 82.79\n \n 80.10\n \n 66.62\n \n 64.36\n \n 87.16\n \n 84.98\n \n 73.90\n \n 71.53\n
\n" }, { "instance_id": "8fedcdae-b77e-42f5-8cf0-9663acf848c0", "table_id_paper": "D18-1396table_4", "caption": "BLEU scores on En-Jp test set. \u201d0\u201d represents the normal translation results, and \u201d1\u201d represents the teacher-forcing translation results.", "row_header_level": 1, "row_headers": [ [ "left" ], [ "right" ] ], "column_header_level": 2, "column_headers": [ [ "left-to-right", "0" ], [ "left-to-right", "1" ], [ "right-to-left", "0" ], [ "right-to-left", "1" ] ], "contents": [ [ "7.90", "9.91", "7.45", "8.95" ], [ "8.70", "11.52", "9.24", "10.59" ] ], "metrics_loc": "column", "metrics_type": "['BLEU', 'BLEU', 'BLEU', 'BLEU']", "target_entity": "['1']", "table_html": "
left-to-right || 0 left-to-right || 1 right-to-left || 0 right-to-left || 1
left 7.90 9.91 7.45 8.95
right 8.70 11.52 9.24 10.59
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D18-1396", "page_no": 6, "dir": "emnlp2018", "description": "Table 4 shows the BLEU score on the En-Jp test set. It can be observed that regardless of decoding direction (i.e., from left-to-right or from right-toleft) and with or without teacher forcing, the accuracy of the right half is always higher than that in the left half. This observation on Japanese is opposite to English, German and Chinese in Section3.1and3.2, and motivates us to investigate the differences between these languages.", "class_sentence": "[1, 1, 2]", "sentences": "['Table 4 shows the BLEU score on the En-Jp test set.', 'It can be observed that regardless of decoding direction (i.e., from left-to-right or from right-toleft) and with or without teacher forcing, the accuracy of the right half is always higher than that in the left half.', 'This observation on Japanese is opposite to English, German and Chinese in Section3.1and3.2, and motivates us to investigate the differences between these languages.']", "header_mention": "[None, ['right'], None]", "valid": 1, "image_id": "D18-1396-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

BLEU scores on En-Jp test set. \u201d0\u201d represents the normal translation results, and \u201d1\u201d represents the teacher-forcing translation results.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n left-to-right || 0left-to-right || 1right-to-left || 0right-to-left || 1
left7.909.917.458.95
right8.7011.529.2410.59
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{BLEU scores on En-Jp test set. \u201d0\u201d represents the normal translation results, and \u201d1\u201d represents the teacher-forcing translation results.}\n\\begin{tabular}{lrrrr}\n\\toprule\n & left-to-right $\\parallel$ 0 & left-to-right $\\parallel$ 1 & right-to-left $\\parallel$ 0 & right-to-left $\\parallel$ 1 \\\\\n\\midrule\nleft & 7.9 & 9.91 & 7.45 & 8.95 \\\\\nright & 8.7 & 11.52 & 9.24 & 10.59 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: BLEU scores on En-Jp test set. \u201d0\u201d represents the normal translation results, and \u201d1\u201d represents the teacher-forcing translation results.\n
\n \n left-to-right || 0\n \n left-to-right || 1\n \n right-to-left || 0\n \n right-to-left || 1\n
\n left\n \n 7.90\n \n 9.91\n \n 7.45\n \n 8.95\n
\n right\n \n 8.70\n \n 11.52\n \n 9.24\n \n 10.59\n
\n" }, { "instance_id": "673f5ced-2927-4c8b-afd6-8ca90fd30167", "table_id_paper": "D18-1400table_3", "caption": "Translation results on the IKEA dataset", "row_header_level": 2, "row_headers": [ [ "Method", "LIUMCVC-Multi" ], [ "Method", "Text-Only NMT" ], [ "Method", "VAG-NMT" ] ], "column_header_level": 2, "column_headers": [ [ "English \u2192 German", "BLEU" ], [ "English \u2192 German", "METEOR" ], [ "English \u2192 French", "BLEU" ], [ "English \u2192 French", "METEOR" ] ], "contents": [ [ "59.9 \u00b1 1.9", "63.8 \u00b1 0.4", "58.4 \u00b1 1.6", "64.6 \u00b1 1.8" ], [ "61.9 \u00b1 0.9", "65.6 \u00b1 0.9", "65.2 \u00b1 0.7", "69.0 \u00b1 0.2" ], [ "63.5 \u00b1 1.2", "65.7 \u00b1 0.1", "65.8 \u00b1 1.2", "68.9 \u00b1 1.4" ] ], "metrics_loc": "column", "metrics_type": "['BLEU', 'METEOR', 'BLEU', 'METEOR']", "target_entity": "['VAG-NMT']", "table_html": "
English \u2192 German || BLEU English \u2192 German || METEOR English \u2192 French || BLEU English \u2192 French || METEOR
Method || LIUMCVC-Multi 59.9 \u00b1 1.9 63.8 \u00b1 0.4 58.4 \u00b1 1.6 64.6 \u00b1 1.8
Method || Text-Only NMT 61.9 \u00b1 0.9 65.6 \u00b1 0.9 65.2 \u00b1 0.7 69.0 \u00b1 0.2
Method || VAG-NMT 63.5 \u00b1 1.2 65.7 \u00b1 0.1 65.8 \u00b1 1.2 68.9 \u00b1 1.4
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D18-1400", "page_no": 6, "dir": "emnlp2018", "description": "We then evaluate all models on the IKEA dataset. Table 3 shows the results. Our VAG-NMT has a higher value in BLEU and a comparable value in METEOR compared to the Text-only NMT baseline. Our VAG-NMT outperforms LIUMCVC's multimodal system by a large margin, which shows that the LIUMCVC's multimodal's good performance on Multi30K does not generalize to this real-world product dataset. The good performance may come from the visual attention mechanism that learns to focus on text segments that are related to the images. Such attention therefore teaches the decoder to apply the visual context to translate those words. This learned attention is especially useful for datasets with long sentences that have irrelevant text information with respect to the image.", "class_sentence": "[2, 1, 1, 1, 2, 2, 2]", "sentences": "['We then evaluate all models on the IKEA dataset.', 'Table 3 shows the results.', 'Our VAG-NMT has a higher value in BLEU and a comparable value in METEOR compared to the Text-only NMT baseline.', \"Our VAG-NMT outperforms LIUMCVC's multimodal system by a large margin, which shows that the LIUMCVC's multimodal's good performance on Multi30K does not generalize to this real-world product dataset.\", 'The good performance may come from the visual attention mechanism that learns to focus on text segments that are related to the images.', 'Such attention therefore teaches the decoder to apply the visual context to translate those words.', 'This learned attention is especially useful for datasets with long sentences that have irrelevant text information with respect to the image.']", "header_mention": "[None, None, ['VAG-NMT', 'Text-Only NMT', 'BLEU', 'METEOR'], ['VAG-NMT', 'LIUMCVC-Multi'], None, None, None]", "valid": 1, "image_id": "D18-1400-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Translation results on the IKEA dataset

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n English \u2192 German || BLEUEnglish \u2192 German || METEOREnglish \u2192 French || BLEUEnglish \u2192 French || METEOR
Method || LIUMCVC-Multi59.9 \u00b1 1.963.8 \u00b1 0.458.4 \u00b1 1.664.6 \u00b1 1.8
Method || Text-Only NMT61.9 \u00b1 0.965.6 \u00b1 0.965.2 \u00b1 0.769.0 \u00b1 0.2
Method || VAG-NMT63.5 \u00b1 1.265.7 \u00b1 0.165.8 \u00b1 1.268.9 \u00b1 1.4
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Translation results on the IKEA dataset}\n\\begin{tabular}{lllll}\n\\toprule\n & English $\\rightarrow$ German $\\parallel$ BLEU & English $\\rightarrow$ German $\\parallel$ METEOR & English $\\rightarrow$ French $\\parallel$ BLEU & English $\\rightarrow$ French $\\parallel$ METEOR \\\\\n\\midrule\nMethod $\\parallel$ LIUMCVC-Multi & 59.9 $\\pm$ 1.9 & 63.8 $\\pm$ 0.4 & 58.4 $\\pm$ 1.6 & 64.6 $\\pm$ 1.8 \\\\\nMethod $\\parallel$ Text-Only NMT & 61.9 $\\pm$ 0.9 & 65.6 $\\pm$ 0.9 & 65.2 $\\pm$ 0.7 & 69.0 $\\pm$ 0.2 \\\\\nMethod $\\parallel$ VAG-NMT & 63.5 $\\pm$ 1.2 & 65.7 $\\pm$ 0.1 & 65.8 $\\pm$ 1.2 & 68.9 $\\pm$ 1.4 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Translation results on the IKEA dataset\n
\n \n English \u2192 German || BLEU\n \n English \u2192 German || METEOR\n \n English \u2192 French || BLEU\n \n English \u2192 French || METEOR\n
\n Method || LIUMCVC-Multi\n \n 59.9 \u00b1 1.9\n \n 63.8 \u00b1 0.4\n \n 58.4 \u00b1 1.6\n \n 64.6 \u00b1 1.8\n
\n Method || Text-Only NMT\n \n 61.9 \u00b1 0.9\n \n 65.6 \u00b1 0.9\n \n 65.2 \u00b1 0.7\n \n 69.0 \u00b1 0.2\n
\n Method || VAG-NMT\n \n 63.5 \u00b1 1.2\n \n 65.7 \u00b1 0.1\n \n 65.8 \u00b1 1.2\n \n 68.9 \u00b1 1.4\n
\n" }, { "instance_id": "bc1039ce-15a2-402a-ab33-c0a670e195ee", "table_id_paper": "D18-1401table_2", "caption": "Performance of our approaches to QA-style sentiment classification in all domains.", "row_header_level": 1, "row_headers": [ [ "SVM" ], [ "LSTM" ], [ "Bi-LSTM" ], [ "Bidirectional-Match" ], [ "AtoQ-Match" ], [ "QtoA-Match" ], [ "Bidirectional-Match QA" ], [ "HMN" ] ], "column_header_level": 2, "column_headers": [ [ "Beauty", "Macro-F1" ], [ "Beauty", "Accuracy" ], [ "Shoe", "Macro-F1" ], [ "Shoe", "Accuracy" ], [ "Electronic", "Macro-F1" ], [ "Electronic", "Accuracy" ] ], "contents": [ [ "0.362", "0.684", "0.381", "0.718", "0.435", "0.691" ], [ "0.499", "0.712", "0.520", "0.754", "0.562", "0.715" ], [ "0.527", "0.719", "0.531", "0.759", "0.574", "0.723" ], [ "0.526", "0.747", "0.557", "0.796", "0.582", "0.741" ], [ "0.543", "0.745", "0.602", "0.792", "0.567", "0.754" ], [ "0.573", "0.751", "0.647", "0.807", "0.608", "0.752" ], [ "0.583", "0.760", "0.666", "0.815", "0.617", "0.764" ], [ "0.598", "0.776", "0.683", "0.827", "0.640", "0.779" ] ], "metrics_loc": "column", "metrics_type": "['Macro-F1', 'Accuracy', 'Macro-F1', 'Accuracy', 'Macro-F1', 'Accuracy']", "target_entity": "['HMN']", "table_html": "
Beauty || Macro-F1 Beauty || Accuracy Shoe || Macro-F1 Shoe || Accuracy Electronic || Macro-F1 Electronic || Accuracy
SVM 0.362 0.684 0.381 0.718 0.435 0.691
LSTM 0.499 0.712 0.520 0.754 0.562 0.715
Bi-LSTM 0.527 0.719 0.531 0.759 0.574 0.723
Bidirectional-Match 0.526 0.747 0.557 0.796 0.582 0.741
AtoQ-Match 0.543 0.745 0.602 0.792 0.567 0.754
QtoA-Match 0.573 0.751 0.647 0.807 0.608 0.752
Bidirectional-Match QA 0.583 0.760 0.666 0.815 0.617 0.764
HMN 0.598 0.776 0.683 0.827 0.640 0.779
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1401", "page_no": 7, "dir": "emnlp2018", "description": "Table 2 summarizes the experimental results of all the approaches above, and we can find that:. (1) All LSTM-based approaches are superior to SVM, indicating the effectiveness of neural network for this task. (2) The proposed approaches, with novel QA contextual representation, outperform the other baseline approaches. (3) When only employing QA bidirectional matching mechanism, Bidirectional-Match QA, which takes the sentence segmentation strategy, consistently outperforms Bidirectional-Match (without sentence segmentation) in all domains. It confirms our hypothesis that sentence segmentation helps to extract the sentiment matching information between the question and answer. (4) When comparing to QA unidirectional matching mechanism, Bidirectional-Match QA, which employs QA bidirectional matching mechanism, performs better than AtoQMatch and QtoA-Match. It confirms our hypothesis that both the question and answer information contribute to sentiment polarity of the QA text pair. (5) Impressively, the proposed approach HMN significantly outperforms all the other approaches in all domains (p-value<0.05 via t-test). It verifies the advantages of both QA bidirectional matching mechanism and selfmatching attention mechanism for this task.", "class_sentence": "[1, 1, 1, 1, 2, 1, 2, 1, 2]", "sentences": "['Table 2 summarizes the experimental results of all the approaches above, and we can find that:.', '(1) All LSTM-based approaches are superior to SVM, indicating the effectiveness of neural network for this task.', '(2) The proposed approaches, with novel QA contextual representation, outperform the other baseline approaches.', '(3) When only employing QA bidirectional matching mechanism, Bidirectional-Match QA, which takes the sentence segmentation strategy, consistently outperforms Bidirectional-Match (without sentence segmentation) in all domains.', 'It confirms our hypothesis that sentence segmentation helps to extract the sentiment matching information between the question and answer.', '(4) When comparing to QA unidirectional matching mechanism, Bidirectional-Match QA, which employs QA bidirectional matching mechanism, performs better than AtoQMatch and QtoA-Match.', 'It confirms our hypothesis that both the question and answer information contribute to sentiment polarity of the QA text pair.', '(5) Impressively, the proposed approach HMN significantly outperforms all the other approaches in all domains (p-value<0.05 via t-test).', 'It verifies the advantages of both QA bidirectional matching mechanism and selfmatching attention mechanism for this task.']", "header_mention": "[None, ['SVM', 'LSTM', 'Bi-LSTM'], ['HMN'], ['Bidirectional-Match', 'Bidirectional-Match QA'], None, ['AtoQ-Match', 'QtoA-Match', 'Bidirectional-Match QA'], None, ['HMN'], ['HMN']]", "valid": 1, "image_id": "D18-1401-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance of our approaches to QA-style sentiment classification in all domains.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Beauty || Macro-F1Beauty || AccuracyShoe || Macro-F1Shoe || AccuracyElectronic || Macro-F1Electronic || Accuracy
SVM0.3620.6840.3810.7180.4350.691
LSTM0.4990.7120.5200.7540.5620.715
Bi-LSTM0.5270.7190.5310.7590.5740.723
Bidirectional-Match0.5260.7470.5570.7960.5820.741
AtoQ-Match0.5430.7450.6020.7920.5670.754
QtoA-Match0.5730.7510.6470.8070.6080.752
Bidirectional-Match QA0.5830.7600.6660.8150.6170.764
HMN0.5980.7760.6830.8270.6400.779
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance of our approaches to QA-style sentiment classification in all domains.}\n\\begin{tabular}{lrrrrrr}\n\\toprule\n & Beauty $\\parallel$ Macro-F1 & Beauty $\\parallel$ Accuracy & Shoe $\\parallel$ Macro-F1 & Shoe $\\parallel$ Accuracy & Electronic $\\parallel$ Macro-F1 & Electronic $\\parallel$ Accuracy \\\\\n\\midrule\nSVM & 0.362 & 0.684 & 0.381 & 0.718 & 0.435 & 0.691 \\\\\nLSTM & 0.499 & 0.712 & 0.52 & 0.754 & 0.562 & 0.715 \\\\\nBi-LSTM & 0.527 & 0.719 & 0.531 & 0.759 & 0.574 & 0.723 \\\\\nBidirectional-Match & 0.526 & 0.747 & 0.557 & 0.796 & 0.582 & 0.741 \\\\\nAtoQ-Match & 0.543 & 0.745 & 0.602 & 0.792 & 0.567 & 0.754 \\\\\nQtoA-Match & 0.573 & 0.751 & 0.647 & 0.807 & 0.608 & 0.752 \\\\\nBidirectional-Match QA & 0.583 & 0.76 & 0.666 & 0.815 & 0.617 & 0.764 \\\\\nHMN & 0.598 & 0.776 & 0.683 & 0.827 & 0.64 & 0.779 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Performance of our approaches to QA-style sentiment classification in all domains.\n
\n \n Beauty || Macro-F1\n \n Beauty || Accuracy\n \n Shoe || Macro-F1\n \n Shoe || Accuracy\n \n Electronic || Macro-F1\n \n Electronic || Accuracy\n
\n SVM\n \n 0.362\n \n 0.684\n \n 0.381\n \n 0.718\n \n 0.435\n \n 0.691\n
\n LSTM\n \n 0.499\n \n 0.712\n \n 0.520\n \n 0.754\n \n 0.562\n \n 0.715\n
\n Bi-LSTM\n \n 0.527\n \n 0.719\n \n 0.531\n \n 0.759\n \n 0.574\n \n 0.723\n
\n Bidirectional-Match\n \n 0.526\n \n 0.747\n \n 0.557\n \n 0.796\n \n 0.582\n \n 0.741\n
\n AtoQ-Match\n \n 0.543\n \n 0.745\n \n 0.602\n \n 0.792\n \n 0.567\n \n 0.754\n
\n QtoA-Match\n \n 0.573\n \n 0.751\n \n 0.647\n \n 0.807\n \n 0.608\n \n 0.752\n
\n Bidirectional-Match QA\n \n 0.583\n \n 0.760\n \n 0.666\n \n 0.815\n \n 0.617\n \n 0.764\n
\n HMN\n \n 0.598\n \n 0.776\n \n 0.683\n \n 0.827\n \n 0.640\n \n 0.779\n
\n" }, { "instance_id": "cd758282-4fc9-4f26-875c-bb2ffcdd0720", "table_id_paper": "D18-1408table_2", "caption": "Performance on SNLI and transfer tasks of various sentence encoders. Test accuracies on SNLI, micro and macro averages of accuracies of dev set on transfer tasks are chosen as evaluation metrics.", "row_header_level": 2, "row_headers": [ [ "Model", "BiLSTM-Max" ], [ "Model", "AdaSent" ], [ "Model", "TBCNN" ], [ "Model", "DiSAN" ], [ "Model", "PSAN" ] ], "column_header_level": 1, "column_headers": [ [ "SNLI" ], [ "Micro" ], [ "Macro" ] ], "contents": [ [ "84.5", "85.2", "83.7" ], [ "83.4", "82.0", "80.9" ], [ "82.1", "81.1", "79.3" ], [ "85.6", "84.7", "83.4" ], [ "86.1", "85.7", "84.5" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy']", "target_entity": "['PSAN']", "table_html": "
SNLI Micro Macro
Model || BiLSTM-max 84.5 85.2 83.7
Model || AdaSent 83.4 82.0 80.9
Model || TBCNN 82.1 81.1 79.3
Model || DiSAN 85.6 84.7 83.4
Model || PSAN 86.1 85.7 84.5
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1408", "page_no": 5, "dir": "emnlp2018", "description": "Experiment results of our model and four baselines are shown in Table 2. Micro and macro accuracies are two composite indicators for evaluating transfer performance of tasks whose metric is classification accuracy. Macro accuracy is the proportion of true results in the population of instances from all tasks. Micro accuracy is the arithmetic mean of dev accuracies for each task. PSAN achieves the state-of-the-art performance with considerably fewer parameters, outperforming a RNN-based model, a CNN-based model, a fully attention-based model and a model that utilize syntactic information. Especially when compared with previous best model BiLSTM-Max, PSAN can outperform their model with only 5% of their parameter numbers, demonstrating the effectiveness of our model at extracting semantically important information from a sentence.", "class_sentence": "[1, 2, 2, 2, 1, 1]", "sentences": "['Experiment results of our model and four baselines are shown in Table 2.', 'Micro and macro accuracies are two composite indicators for evaluating transfer performance of tasks whose metric is classification accuracy.', 'Macro accuracy is the proportion of true results in the population of instances from all tasks.', 'Micro accuracy is the arithmetic mean of dev accuracies for each task.', 'PSAN achieves the state-of-the-art performance with considerably fewer parameters, outperforming a RNN-based model, a CNN-based model, a fully attention-based model and a model that utilize syntactic information.', 'Especially when compared with previous best model BiLSTM-Max, PSAN can outperform their model with only 5% of their parameter numbers, demonstrating the effectiveness of our model at extracting semantically important information from a sentence.']", "header_mention": "[['Model'], ['Micro', 'Macro'], ['Macro'], ['Micro'], ['PSAN'], ['PSAN', 'BiLSTM-Max']]", "valid": 1, "image_id": "D18-1408-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance on SNLI and transfer tasks of various sentence encoders. Test accuracies on SNLI, micro and macro averages of accuracies of dev set on transfer tasks are chosen as evaluation metrics.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n SNLIMicroMacro
Model || BiLSTM-max84.585.283.7
Model || AdaSent83.482.080.9
Model || TBCNN82.181.179.3
Model || DiSAN85.684.783.4
Model || PSAN86.185.784.5
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance on SNLI and transfer tasks of various sentence encoders. Test accuracies on SNLI, micro and macro averages of accuracies of dev set on transfer tasks are chosen as evaluation metrics.}\n\\begin{tabular}{lrrr}\n\\toprule\n & SNLI & Micro & Macro \\\\\n\\midrule\nModel $\\parallel$ BiLSTM-max & 84.5 & 85.2 & 83.7 \\\\\nModel $\\parallel$ AdaSent & 83.4 & 82 & 80.9 \\\\\nModel $\\parallel$ TBCNN & 82.1 & 81.1 & 79.3 \\\\\nModel $\\parallel$ DiSAN & 85.6 & 84.7 & 83.4 \\\\\nModel $\\parallel$ PSAN & 86.1 & 85.7 & 84.5 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Performance on SNLI and transfer tasks of various sentence encoders. Test accuracies on SNLI, micro and macro averages of accuracies of dev set on transfer tasks are chosen as evaluation metrics.\n
\n \n SNLI\n \n Micro\n \n Macro\n
\n Model || BiLSTM-max\n \n 84.5\n \n 85.2\n \n 83.7\n
\n Model || AdaSent\n \n 83.4\n \n 82.0\n \n 80.9\n
\n Model || TBCNN\n \n 82.1\n \n 81.1\n \n 79.3\n
\n Model || DiSAN\n \n 85.6\n \n 84.7\n \n 83.4\n
\n Model || PSAN\n \n 86.1\n \n 85.7\n \n 84.5\n
\n" }, { "instance_id": "177a4980-4212-42f3-ad68-6cb6164251f3", "table_id_paper": "D18-1408table_3", "caption": "Transfer test results for our model and various baselines. Classification accuracy is chosen as evaluation metric for datasets including MR, CR, SUBJ, MPQA, SST, TREC and SICK-E.", "row_header_level": 2, "row_headers": [ [ "Model", "BiLSTM-max" ], [ "Model", "AdaSent" ], [ "Model", "TBCNN" ], [ "Model", "DiSAN" ], [ "Model", "PSAN" ] ], "column_header_level": 1, "column_headers": [ [ "MR" ], [ "CR" ], [ "SUBJ" ], [ "MPQA" ], [ "SST" ], [ "TREC" ] ], "contents": [ [ "79.9", "84.6", "92.1", "89.8", "83.3", "88.7" ], [ "77.0", "82.0", "89.9", "87.2", "82.3", "85.6" ], [ "75.4", "81.6", "89.1", "85.9", "79.4", "83.7" ], [ "79.7", "84.1", "92.2", "89.5", "82.9", "88.3" ], [ "80.0", "84.2", "91.9", "89.9", "83.8", "89.1" ] ], "metrics_loc": "column", "metrics_type": "['Accuracy', 'Accuracy', 'Accuracy', 'Accuracy', 'Accuracy', 'Accuracy']", "target_entity": "['PSAN']", "table_html": "
MR CR SUBJ MPQA SST TREC
Model || BiLSTM-max 79.9 84.6 92.1 89.8 83.3 88.7
Model || AdaSent 77.0 82.0 89.9 87.2 82.3 85.6
Model || TBCNN 75.4 81.6 89.1 85.9 79.4 83.7
Model || DiSAN 79.7 84.1 92.2 89.5 82.9 88.3
Model || PSAN 80.0 84.2 91.9 89.9 83.8 89.1
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D18-1408", "page_no": 6, "dir": "emnlp2018", "description": "In Table 3, we compare our model with baseline sentence encoders in each transfer task. PSAN can consistently outperform the baselines in almost every task considered.", "class_sentence": "[1, 1]", "sentences": "['In Table 3, we compare our model with baseline sentence encoders in each transfer task.', 'PSAN can consistently outperform the baselines in almost every task considered.']", "header_mention": "[None, ['PSAN']]", "valid": 1, "image_id": "D18-1408-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Transfer test results for our model and various baselines. Classification accuracy is chosen as evaluation metric for datasets including MR, CR, SUBJ, MPQA, SST, TREC and SICK-E.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n MRCRSUBJMPQASSTTREC
Model || BiLSTM-max79.984.692.189.883.388.7
Model || AdaSent77.082.089.987.282.385.6
Model || TBCNN75.481.689.185.979.483.7
Model || DiSAN79.784.192.289.582.988.3
Model || PSAN80.084.291.989.983.889.1
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Transfer test results for our model and various baselines. Classification accuracy is chosen as evaluation metric for datasets including MR, CR, SUBJ, MPQA, SST, TREC and SICK-E.}\n\\begin{tabular}{lrrrrrr}\n\\toprule\n & MR & CR & SUBJ & MPQA & SST & TREC \\\\\n\\midrule\nModel $\\parallel$ BiLSTM-max & 79.9 & 84.6 & 92.1 & 89.8 & 83.3 & 88.7 \\\\\nModel $\\parallel$ AdaSent & 77 & 82 & 89.9 & 87.2 & 82.3 & 85.6 \\\\\nModel $\\parallel$ TBCNN & 75.4 & 81.6 & 89.1 & 85.9 & 79.4 & 83.7 \\\\\nModel $\\parallel$ DiSAN & 79.7 & 84.1 & 92.2 & 89.5 & 82.9 & 88.3 \\\\\nModel $\\parallel$ PSAN & 80 & 84.2 & 91.9 & 89.9 & 83.8 & 89.1 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Transfer test results for our model and various baselines. Classification accuracy is chosen as evaluation metric for datasets including MR, CR, SUBJ, MPQA, SST, TREC and SICK-E.\n
\n \n MR\n \n CR\n \n SUBJ\n \n MPQA\n \n SST\n \n TREC\n
\n Model || BiLSTM-max\n \n 79.9\n \n 84.6\n \n 92.1\n \n 89.8\n \n 83.3\n \n 88.7\n
\n Model || AdaSent\n \n 77.0\n \n 82.0\n \n 89.9\n \n 87.2\n \n 82.3\n \n 85.6\n
\n Model || TBCNN\n \n 75.4\n \n 81.6\n \n 89.1\n \n 85.9\n \n 79.4\n \n 83.7\n
\n Model || DiSAN\n \n 79.7\n \n 84.1\n \n 92.2\n \n 89.5\n \n 82.9\n \n 88.3\n
\n Model || PSAN\n \n 80.0\n \n 84.2\n \n 91.9\n \n 89.9\n \n 83.8\n \n 89.1\n
\n" }, { "instance_id": "fd11d89d-d9b6-4032-90be-ff8fda519117", "table_id_paper": "D18-1417table_2", "caption": "Results of independent training for intent detection in terms of error rate.", "row_header_level": 2, "row_headers": [ [ "Methods", "Recursive NN (Guo et al. 2014)" ], [ "Methods", "Boosting (Tur et al. 2010)" ], [ "Methods", "Boosting + Simplified sentences(Tur et al. 2011)" ], [ "Methods", "Attention Enc-Dec (Liu and Lane 2016a)" ], [ "Methods", "Our Model" ] ], "column_header_level": 1, "column_headers": [ [ "Error(%)" ] ], "contents": [ [ "4.60" ], [ "4.38" ], [ "3.02" ], [ "2.02" ], [ "2.69" ] ], "metrics_loc": "column", "metrics_type": "['Error(%)']", "target_entity": "['Our Model']", "table_html": "
Error(%)
Methods || Recursive NN (Guo et al. 2014) 4.60
Methods || Boosting (Tur et al. 2010) 4.38
Methods || Boosting + Simplified sentences(Tur et al. 2011) 3.02
Methods || Attention Enc-Dec (Liu and Lane 2016a) 2.02
Methods || Our Model 2.69
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1417", "page_no": 7, "dir": "emnlp2018", "description": "4.4 Independent Learning. The results of separate training for slot filling and intent detection are reported in Table 1 and Table 2 respectively. On the independent slot filling task, we fixed the intent information as the ground truth labels in the dataset. But on the independent intent detection task, there is no interaction with slot labels. Table 2 compares the performance of our proposed model to previously reported results on intent detection task. Our model gives good performance in terms of classification error rate, but not as good as Attention Encoder-Decoder (with aligned inputs) method (Liu and Lane, 2016a). As their published state-of-the-art result described in (Liu and Lane, 2016a), their attention-based model is based on word-level embeddings. While in our model, we introduce character-level embeddings to improve the performance of joint learning. But independent learning for intent classification aims at capturing the global information of an utterance, not caring much about the details of specific word. The character-level embeddings introduced in our model bring very little hurt to independent learning of intent detection, as a trade-off in performance between both criterion.", "class_sentence": "[2, 1, 2, 0, 1, 1, 2, 2, 2, 2]", "sentences": "['4.4 Independent Learning.', 'The results of separate training for slot filling and intent detection are reported in Table 1 and Table 2 respectively.', 'On the independent slot filling task, we fixed the intent information as the ground truth labels in the dataset.', 'But on the independent intent detection task, there is no interaction with slot labels.', 'Table 2 compares the performance of our proposed model to previously reported results on intent detection task.', 'Our model gives good performance in terms of classification error rate, but not as good as Attention Encoder-Decoder (with aligned inputs) method (Liu and Lane, 2016a).', 'As their published state-of-the-art result described in (Liu and Lane, 2016a), their attention-based model is based on word-level embeddings.', 'While in our model, we introduce character-level embeddings to improve the performance of joint learning.', 'But independent learning for intent classification aims at capturing the global information of an utterance, not caring much about the details of specific word.', 'The character-level embeddings introduced in our model bring very little hurt to independent learning of intent detection, as a trade-off in performance between both criterion.']", "header_mention": "[None, None, None, None, ['Our Model'], ['Our Model', 'Attention Enc-Dec (Liu and Lane 2016a)', 'Error(%)'], ['Attention Enc-Dec (Liu and Lane 2016a)'], ['Our Model'], None, ['Our Model']]", "valid": 1, "image_id": "D18-1417-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results of independent training for intent detection in terms of error rate.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Error(%)
Methods || Recursive NN (Guo et al. 2014)4.60
Methods || Boosting (Tur et al. 2010)4.38
Methods || Boosting + Simplified sentences(Tur et al. 2011)3.02
Methods || Attention Enc-Dec (Liu and Lane 2016a)2.02
Methods || Our Model2.69
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results of independent training for intent detection in terms of error rate.}\n\\begin{tabular}{lr}\n\\toprule\n & Error(\\%) \\\\\n\\midrule\nMethods $\\parallel$ Recursive NN (Guo et al. 2014) & 4.6 \\\\\nMethods $\\parallel$ Boosting (Tur et al. 2010) & 4.38 \\\\\nMethods $\\parallel$ Boosting + Simplified sentences(Tur et al. 2011) & 3.02 \\\\\nMethods $\\parallel$ Attention Enc-Dec (Liu and Lane 2016a) & 2.02 \\\\\nMethods $\\parallel$ Our Model & 2.69 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Results of independent training for intent detection in terms of error rate.\n
\n \n Error(%)\n
\n Methods || Recursive NN (Guo et al. 2014)\n \n 4.60\n
\n Methods || Boosting (Tur et al. 2010)\n \n 4.38\n
\n Methods || Boosting + Simplified sentences(Tur et al. 2011)\n \n 3.02\n
\n Methods || Attention Enc-Dec (Liu and Lane 2016a)\n \n 2.02\n
\n Methods || Our Model\n \n 2.69\n
\n" }, { "instance_id": "82b94266-f6b9-4aa8-9163-5b443abed051", "table_id_paper": "D18-1417table_3", "caption": "Results of joint training for slot filling and intent detection.", "row_header_level": 2, "row_headers": [ [ "Methods", "Recursive NN (Guo et al.2014)" ], [ "Methods", "Recursive NN+Viterbi(Guo et al. 2014)" ], [ "Methods", "Attention Enc-Dec (Liu and Lane 2016a)" ], [ "Methods", "Attention BiRNN (Liu and Lane 2016a)" ], [ "Methods", "Our Model" ] ], "column_header_level": 1, "column_headers": [ [ "F1" ], [ "Error(%)" ] ], "contents": [ [ "93.22", "4.60" ], [ "93.96", "4.60" ], [ "95.87", "1.57" ], [ "95.98", "1.79" ], [ "96.52", "1.23" ] ], "metrics_loc": "column", "metrics_type": "['F1', 'Error(%)']", "target_entity": "['Our Model']", "table_html": "
F1 Error(%)
Methods || Recursive NN (Guo et al.2014) 93.22 4.60
Methods || Recursive NN+Viterbi(Guo et al. 2014) 93.96 4.60
Methods || Attention Enc-Dec (Liu and Lane 2016a) 95.87 1.57
Methods || Attention BiRNN (Liu and Lane 2016a) 95.98 1.79
Methods || Our Model 96.52 1.23
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D18-1417", "page_no": 7, "dir": "emnlp2018", "description": "Table 3 compares our joint model with reported results from previous works. We can see that our model achieves state-of-the-art results and outperforms previous best result by 0.54% in terms of F1-score on slot filling, and by 0.34% in terms of error rate on intent detection. This improvement is statistically significant. Besides, the joint learning achieves better results than separate learning. It can be interpreted that the two tasks are highly correlated and boost the performance each other. The slot filling task enables the model to learn more meaningful representations which give more supervisory signals for the learning of shared parameters. Similarly, intent is also useful to determine the slot label.", "class_sentence": "[1, 1, 2, 1, 2, 2, 2]", "sentences": "['Table 3 compares our joint model with reported results from previous works.', 'We can see that our model achieves state-of-the-art results and outperforms previous best result by 0.54% in terms of F1-score on slot filling, and by 0.34% in terms of error rate on intent detection.', 'This improvement is statistically significant.', 'Besides, the joint learning achieves better results than separate learning.', 'It can be interpreted that the two tasks are highly correlated and boost the performance each other.', 'The slot filling task enables the model to learn more meaningful representations which give more supervisory signals for the learning of shared parameters.', 'Similarly, intent is also useful to determine the slot label.']", "header_mention": "[['Our Model'], ['Our Model', 'F1', 'Error(%)'], ['Our Model'], None, None, None, None]", "valid": 1, "image_id": "D18-1417-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results of joint training for slot filling and intent detection.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n F1Error(%)
Methods || Recursive NN (Guo et al.2014)93.224.60
Methods || Recursive NN+Viterbi(Guo et al. 2014)93.964.60
Methods || Attention Enc-Dec (Liu and Lane 2016a)95.871.57
Methods || Attention BiRNN (Liu and Lane 2016a)95.981.79
Methods || Our Model96.521.23
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results of joint training for slot filling and intent detection.}\n\\begin{tabular}{lrr}\n\\toprule\n & F1 & Error(\\%) \\\\\n\\midrule\nMethods $\\parallel$ Recursive NN (Guo et al.2014) & 93.22 & 4.6 \\\\\nMethods $\\parallel$ Recursive NN+Viterbi(Guo et al. 2014) & 93.96 & 4.6 \\\\\nMethods $\\parallel$ Attention Enc-Dec (Liu and Lane 2016a) & 95.87 & 1.57 \\\\\nMethods $\\parallel$ Attention BiRNN (Liu and Lane 2016a) & 95.98 & 1.79 \\\\\nMethods $\\parallel$ Our Model & 96.52 & 1.23 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Results of joint training for slot filling and intent detection.\n
\n \n F1\n \n Error(%)\n
\n Methods || Recursive NN (Guo et al.2014)\n \n 93.22\n \n 4.60\n
\n Methods || Recursive NN+Viterbi(Guo et al. 2014)\n \n 93.96\n \n 4.60\n
\n Methods || Attention Enc-Dec (Liu and Lane 2016a)\n \n 95.87\n \n 1.57\n
\n Methods || Attention BiRNN (Liu and Lane 2016a)\n \n 95.98\n \n 1.79\n
\n Methods || Our Model\n \n 96.52\n \n 1.23\n
\n" }, { "instance_id": "a2002ec8-ca3b-4266-8214-78f3deb657d4", "table_id_paper": "D18-1421table_3", "caption": "Performances on Twitter corpus.", "row_header_level": 2, "row_headers": [ [ "Models", "Seq2Seq" ], [ "Models", "Residual LSTM" ], [ "Models", "Pointer-generator" ], [ "Models", "RL-ROUGE" ], [ "Models", "RbM-SL (ours)" ], [ "Models", "RbM-IRL (ours)" ] ], "column_header_level": 2, "column_headers": [ [ "Twitter", "ROUGE-1" ], [ "Twitter", "ROUGE-2" ], [ "Twitter", "BLEU" ], [ "Twitter", "METEOR" ] ], "contents": [ [ "30.43", "14.61", "30.54", "12.80" ], [ "32.50", "16.86", "33.90", "13.65" ], [ "38.31", "21.22", "40.37", "17.62" ], [ "40.16", "22.99", "42.73", "18.89" ], [ "41.87", "24.23", "44.67", "19.97" ], [ "42.15", "24.73", "45.74", "20.18" ] ], "metrics_loc": "column", "metrics_type": "['ROUGE-1', 'ROUGE-2', 'BLEU', 'METEOR']", "target_entity": "['RbM-SL (ours)', 'RbM-IRL (ours)']", "table_html": "
Twitter || ROUGE-1 Twitter || ROUGE-2 Twitter || BLEU Twitter || METEOR
Models || Seq2Seq 30.43 14.61 30.54 12.80
Models || Residual LSTM 32.50 16.86 33.90 13.65
Models || Pointer-generator 38.31 21.22 40.37 17.62
Models || RL-ROUGE 40.16 22.99 42.73 18.89
Models || RbM-SL (ours) 41.87 24.23 44.67 19.97
Models || RbM-IRL (ours) 42.15 24.73 45.74 20.18
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D18-1421", "page_no": 7, "dir": "emnlp2018", "description": "Table 3 shows the performances on the Twitter corpus. Our models again outperform the baselines in terms of all the evaluation measures. Note that RbM-IRL performs better than RbM-SL in this case. The reason might be that the evaluator of RbM-SL might not be effectively trained with the relatively small dataset, while RbM-IRL can leverage its advantage in learning of the evaluator with less data.", "class_sentence": "[1, 1, 1, 2]", "sentences": "['Table 3 shows the performances on the Twitter corpus.', 'Our models again outperform the baselines in terms of all the evaluation measures.', 'Note that RbM-IRL performs better than RbM-SL in this case.', 'The reason might be that the evaluator of RbM-SL might not be effectively trained with the relatively small dataset, while RbM-IRL can leverage its advantage in learning of the evaluator with less data.']", "header_mention": "[None, ['RbM-SL (ours)', 'RbM-IRL (ours)'], ['RbM-IRL (ours)', 'RbM-SL (ours)'], ['RbM-SL (ours)', 'RbM-IRL (ours)']]", "valid": 1, "image_id": "D18-1421-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performances on Twitter corpus.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Twitter || ROUGE-1Twitter || ROUGE-2Twitter || BLEUTwitter || METEOR
Models || Seq2Seq30.4314.6130.5412.80
Models || Residual LSTM32.5016.8633.9013.65
Models || Pointer-generator38.3121.2240.3717.62
Models || RL-ROUGE40.1622.9942.7318.89
Models || RbM-SL (ours)41.8724.2344.6719.97
Models || RbM-IRL (ours)42.1524.7345.7420.18
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performances on Twitter corpus.}\n\\begin{tabular}{lrrrr}\n\\toprule\n & Twitter $\\parallel$ ROUGE-1 & Twitter $\\parallel$ ROUGE-2 & Twitter $\\parallel$ BLEU & Twitter $\\parallel$ METEOR \\\\\n\\midrule\nModels $\\parallel$ Seq2Seq & 30.43 & 14.61 & 30.54 & 12.8 \\\\\nModels $\\parallel$ Residual LSTM & 32.5 & 16.86 & 33.9 & 13.65 \\\\\nModels $\\parallel$ Pointer-generator & 38.31 & 21.22 & 40.37 & 17.62 \\\\\nModels $\\parallel$ RL-ROUGE & 40.16 & 22.99 & 42.73 & 18.89 \\\\\nModels $\\parallel$ RbM-SL (ours) & 41.87 & 24.23 & 44.67 & 19.97 \\\\\nModels $\\parallel$ RbM-IRL (ours) & 42.15 & 24.73 & 45.74 & 20.18 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Performances on Twitter corpus.\n
\n \n Twitter || ROUGE-1\n \n Twitter || ROUGE-2\n \n Twitter || BLEU\n \n Twitter || METEOR\n
\n Models || Seq2Seq\n \n 30.43\n \n 14.61\n \n 30.54\n \n 12.80\n
\n Models || Residual LSTM\n \n 32.50\n \n 16.86\n \n 33.90\n \n 13.65\n
\n Models || Pointer-generator\n \n 38.31\n \n 21.22\n \n 40.37\n \n 17.62\n
\n Models || RL-ROUGE\n \n 40.16\n \n 22.99\n \n 42.73\n \n 18.89\n
\n Models || RbM-SL (ours)\n \n 41.87\n \n 24.23\n \n 44.67\n \n 19.97\n
\n Models || RbM-IRL (ours)\n \n 42.15\n \n 24.73\n \n 45.74\n \n 20.18\n
\n" }, { "instance_id": "817a85e7-806a-4894-862d-6d4eedf9b374", "table_id_paper": "D18-1448table_2", "caption": "User satisfaction test results. In total, we use the strategy mentioned in Section 4.2 to evaluate 400 randomly selected source news pages. Overall denotes the average score on these 400 samples.", "row_header_level": 2, "row_headers": [ [ "Format", "Text" ], [ "Format", "Pictorial" ] ], "column_header_level": 1, "column_headers": [ [ "AnnotatorA" ], [ "AnnotatorB" ], [ "Overall" ] ], "contents": [ [ "3.67", "3.75", "3.71" ], [ "4.14", "4.20", "4.17" ] ], "metrics_loc": "column", "metrics_type": "['satisfaction', 'satisfaction', 'satisfaction']", "target_entity": "['Pictorial']", "table_html": "
AnnotatorA AnnotatorB Overall
Format || Text 3.67 3.75 3.71
Format || Pictorial 4.14 4.20 4.17
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D18-1448", "page_no": 6, "dir": "emnlp2018", "description": "Table 2 shows our results for user satisfaction test. User ratings of pictorial summaries are 12.4% higher than text summaries. It shows that users prefer this way of presenting information. It also confirms our motivation for MSMO.", "class_sentence": "[1, 1, 1, 2]", "sentences": "['Table 2 shows our results for user satisfaction test.', 'User ratings of pictorial summaries are 12.4% higher than text summaries.', 'It shows that users prefer this way of presenting information.', 'It also confirms our motivation for MSMO.']", "header_mention": "[None, ['Text', 'Pictorial'], ['Pictorial'], None]", "valid": 1, "image_id": "D18-1448-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

User satisfaction test results. In total, we use the strategy mentioned in Section 4.2 to evaluate 400 randomly selected source news pages. Overall denotes the average score on these 400 samples.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n AnnotatorAAnnotatorBOverall
Format || Text3.673.753.71
Format || Pictorial4.144.204.17
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{User satisfaction test results. In total, we use the strategy mentioned in Section 4.2 to evaluate 400 randomly selected source news pages. Overall denotes the average score on these 400 samples.}\n\\begin{tabular}{lrrr}\n\\toprule\n & AnnotatorA & AnnotatorB & Overall \\\\\n\\midrule\nFormat $\\parallel$ Text & 3.67 & 3.75 & 3.71 \\\\\nFormat $\\parallel$ Pictorial & 4.14 & 4.2 & 4.17 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: User satisfaction test results. In total, we use the strategy mentioned in Section 4.2 to evaluate 400 randomly selected source news pages. Overall denotes the average score on these 400 samples.\n
\n \n AnnotatorA\n \n AnnotatorB\n \n Overall\n
\n Format || Text\n \n 3.67\n \n 3.75\n \n 3.71\n
\n Format || Pictorial\n \n 4.14\n \n 4.20\n \n 4.17\n
\n" }, { "instance_id": "1d384e68-28fa-4b3b-ab1d-fd4ad75f45d7", "table_id_paper": "D18-1486table_3", "caption": "Multi-task results of MCapsNet. In column Avg.\u25b3, we use BiLSTM as baseline and calculate the average improvements over it.", "row_header_level": 1, "row_headers": [ [ "BiLSTM" ], [ "MT-GRNN" ], [ "MT-RNN" ], [ "MT-DNN" ], [ "MT-CNN" ], [ "CapsNet-1" ], [ "CapsNet-2" ], [ "MCapsNet" ] ], "column_header_level": 2, "column_headers": [ [ "Dataset", "MR" ], [ "Dataset", "SST-1" ], [ "Dataset", "SST-2" ], [ "Dataset", "Subj" ], [ "Dataset", "TREC" ], [ "Dataset", "AG\u2019s" ], [ "Dataset", "Avg.\u25b3" ] ], "contents": [ [ "79.3", "46.2", "83.2", "90.5", "89.6", "88.2", "+0" ], [ "-", "49.2", "87.7", "89.3", "93.8", "-", "+2.6" ], [ "-", "49.6", "87.9", "94.1", "91.8", "-", "+3.5" ], [ "82.1", "48.1", "87.3", "93.9", "92.2", "91.8", "+2.9" ], [ "81.6", "49.0", "86.9", "93.6", "91.8", "91.9", "+3.0" ], [ "81.5", "48.1", "86.4", "93.3", "91.8", "91.1", "+2.5" ], [ "82.4", "48.7", "87.8", "93.6", "92.9", "92.3", "+3.3" ], [ "83.5", "49.7", "88.6", "94.5", "94.2", "93.8", "+4.6" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy']", "target_entity": "['MCapsNet']", "table_html": "
Dataset || MR Dataset || SST-1 Dataset || SST-2 Dataset || Subj Dataset || TREC Dataset || AG\u2019s Dataset || Avg.\u25b3
BiLSTM 79.3 46.2 83.2 90.5 89.6 88.2 +0
MT-GRNN - 49.2 87.7 89.3 93.8 - +2.6
MT-RNN - 49.6 87.9 94.1 91.8 - +3.5
MT-DNN 82.1 48.1 87.3 93.9 92.2 91.8 +2.9
MT-CNN 81.6 49.0 86.9 93.6 91.8 91.9 +3.0
CapsNet-1 81.5 48.1 86.4 93.3 91.8 91.1 +2.5
CapsNet-2 82.4 48.7 87.8 93.6 92.9 92.3 +3.3
MCapsNet 83.5 49.7 88.6 94.5 94.2 93.8 +4.6
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D18-1486", "page_no": 7, "dir": "emnlp2018", "description": "Multi-Task Performance. We simultaneously train our model McapsNet on six tasks in Table 1 and compare it with singletask scenario (Table 3). We can see that our multitask architecture clearly improves the performance over the single task models, which demonstrates the benefits of our multi-task architecture. As Table 3 shows, MCapsNet also outperforms the state-of-the-art multi-task learning models by at least 1.1%. This shows the advantages of our task routing algorithm, which can cluster the features for each task, instead of freely sharing the features among tasks.", "class_sentence": "[2, 1, 2, 1, 2]", "sentences": "['Multi-Task Performance.', 'We simultaneously train our model McapsNet on six tasks in Table 1 and compare it with singletask scenario (Table 3).', 'We can see that our multitask architecture clearly improves the performance over the single task models, which demonstrates the benefits of our multi-task architecture.', 'As Table 3 shows, MCapsNet also outperforms the state-of-the-art multi-task learning models by at least 1.1%.', 'This shows the advantages of our task routing algorithm, which can cluster the features for each task, instead of freely sharing the features among tasks.']", "header_mention": "[None, ['MCapsNet'], ['MCapsNet'], ['MCapsNet', 'MT-GRNN', 'MT-RNN', 'MT-DNN', 'MT-CNN'], ['MCapsNet']]", "valid": 1, "image_id": "D18-1486-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Multi-task results of MCapsNet. In column Avg.\u25b3, we use BiLSTM as baseline and calculate the average improvements over it.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Dataset || MRDataset || SST-1Dataset || SST-2Dataset || SubjDataset || TRECDataset || AG\u2019sDataset || Avg.\u25b3
BiLSTM79.346.283.290.589.688.2+0
MT-GRNN-49.287.789.393.8-+2.6
MT-RNN-49.687.994.191.8-+3.5
MT-DNN82.148.187.393.992.291.8+2.9
MT-CNN81.649.086.993.691.891.9+3.0
CapsNet-181.548.186.493.391.891.1+2.5
CapsNet-282.448.787.893.692.992.3+3.3
MCapsNet83.549.788.694.594.293.8+4.6
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Multi-task results of MCapsNet. In column Avg.$\\bigtriangleup$, we use BiLSTM as baseline and calculate the average improvements over it.}\n\\begin{tabular}{llrrrrlr}\n\\toprule\n & Dataset $\\parallel$ MR & Dataset $\\parallel$ SST-1 & Dataset $\\parallel$ SST-2 & Dataset $\\parallel$ Subj & Dataset $\\parallel$ TREC & Dataset $\\parallel$ AG\u2019s & Dataset $\\parallel$ Avg.$\\bigtriangleup$ \\\\\n\\midrule\nBiLSTM & 79.3 & 46.2 & 83.2 & 90.5 & 89.6 & 88.2 & 0 \\\\\nMT-GRNN & - & 49.2 & 87.7 & 89.3 & 93.8 & - & 2.6 \\\\\nMT-RNN & - & 49.6 & 87.9 & 94.1 & 91.8 & - & 3.5 \\\\\nMT-DNN & 82.1 & 48.1 & 87.3 & 93.9 & 92.2 & 91.8 & 2.9 \\\\\nMT-CNN & 81.6 & 49 & 86.9 & 93.6 & 91.8 & 91.9 & 3 \\\\\nCapsNet-1 & 81.5 & 48.1 & 86.4 & 93.3 & 91.8 & 91.1 & 2.5 \\\\\nCapsNet-2 & 82.4 & 48.7 & 87.8 & 93.6 & 92.9 & 92.3 & 3.3 \\\\\nMCapsNet & 83.5 & 49.7 & 88.6 & 94.5 & 94.2 & 93.8 & 4.6 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Multi-task results of MCapsNet. In column Avg.\u25b3, we use BiLSTM as baseline and calculate the average improvements over it.\n
\n \n Dataset || MR\n \n Dataset || SST-1\n \n Dataset || SST-2\n \n Dataset || Subj\n \n Dataset || TREC\n \n Dataset || AG\u2019s\n \n Dataset || Avg.\u25b3\n
\n BiLSTM\n \n 79.3\n \n 46.2\n \n 83.2\n \n 90.5\n \n 89.6\n \n 88.2\n \n +0\n
\n MT-GRNN\n \n -\n \n 49.2\n \n 87.7\n \n 89.3\n \n 93.8\n \n -\n \n +2.6\n
\n MT-RNN\n \n -\n \n 49.6\n \n 87.9\n \n 94.1\n \n 91.8\n \n -\n \n +3.5\n
\n MT-DNN\n \n 82.1\n \n 48.1\n \n 87.3\n \n 93.9\n \n 92.2\n \n 91.8\n \n +2.9\n
\n MT-CNN\n \n 81.6\n \n 49.0\n \n 86.9\n \n 93.6\n \n 91.8\n \n 91.9\n \n +3.0\n
\n CapsNet-1\n \n 81.5\n \n 48.1\n \n 86.4\n \n 93.3\n \n 91.8\n \n 91.1\n \n +2.5\n
\n CapsNet-2\n \n 82.4\n \n 48.7\n \n 87.8\n \n 93.6\n \n 92.9\n \n 92.3\n \n +3.3\n
\n MCapsNet\n \n 83.5\n \n 49.7\n \n 88.6\n \n 94.5\n \n 94.2\n \n 93.8\n \n +4.6\n
\n" }, { "instance_id": "0e846762-42c2-4f36-8f27-627580d7b6b4", "table_id_paper": "D18-1533table_1", "caption": "Performance (%CERR). Slash-separated pairs denote FSTs incapable/capable of inserting word boundaries, respectively; see Section 4. The -KN suffix denotes Kneser-Ney smoothing. The data from Section 3.2 is used for evaluating the modern-orthography language model perplexity, and \u201cCorpus\u201d evaluates test-set transliteration performance from the synthetic missionary text back to the original modern text.", "row_header_level": 2, "row_headers": [ [ "Transliteration model", "FST-(C/Cwb)-7GRAM-KN" ], [ "Transliteration model", "FST-(C/Cwb)-9GRAM-KN" ], [ "Transliteration model", "FST-(C/Cwb)-11GRAM-KN" ], [ "Transliteration model", "FST-RNNLM-(C/Cwb)" ] ], "column_header_level": 2, "column_headers": [ [ "LM perplexity", "Valid." ], [ "LM perplexity", "Test" ], [ "Transliteration performance (%CERR)", "Corpus" ], [ "Transliteration performance (%CERR)", "Newspaper 1" ], [ "Transliteration performance (%CERR)", "Newspaper 2" ] ], "contents": [ [ "3.07", "3.13", "27.3%", "50.1% / 38.7%", "52.0% / 47.5%" ], [ "2.95", "3.02", "26.6%", "50.7% / 39.3%", "52.5% / 47.2%" ], [ "2.94", "3.02", "27.8%", "53.9% / 41.3%", "54.1% / 48.7%" ], [ "2.65", "2.69", "16.3%", "47.2% / 34.3%", "49.8% / 41.2%" ] ], "metrics_loc": "column", "metrics_type": "['LM perplexity', 'LM perplexity', 'Transliteration performance (%CERR)', 'Transliteration performance (%CERR)', 'Transliteration performance (%CERR)']", "target_entity": "['FST-(C/Cwb)-7GRAM-KN', 'FST-(C/Cwb)-9GRAM-KN', 'FST-(C/Cwb)-11GRAM-KN', 'FST-RNNLM-(C/Cwb)']", "table_html": "
LM perplexity || Valid. LM perplexity || Test Transliteration performance (%CERR) || Corpus Transliteration performance (%CERR) || Newspaper 1 Transliteration performance (%CERR) || Newspaper 2
Transliteration model || FST-(C/Cwb)-7GRAM-KN 3.07 3.13 27.3% 50.1% / 38.7% 52.0% / 47.5%
Transliteration model || FST-(C/Cwb)-9GRAM-KN 2.95 3.02 26.6% 50.7% / 39.3% 52.5% / 47.2%
Transliteration model || FST-(C/Cwb)-11GRAM-KN 2.94 3.02 27.8% 53.9% / 41.3% 54.1% / 48.7%
Transliteration model || FST-RNNLM-(C/Cwb) 2.65 2.69 16.3% 47.2% / 34.3% 49.8% / 41.2%
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "D18-1533", "page_no": 5, "dir": "emnlp2018", "description": "Table 1 reports the results of the approaches we described in Section 4. Out of the Kneser-Ney n-gram models, we found that the FST-C-9GRAMKN and the version modelling word boundaries (FST-Cwb-9GRAM-KN) to perform best on the synthetic parallel corpus and newspapers, respectively. Cwb was not applied to the synthetic parallel corpus as we did not model word splitting. The hybrid models (FST-RNNLM) outperformed all FST-only approaches. Thus, we found the wordsplitting hybrid model (FST-RNNLM-Cwb) to be the best performing model overall (Table 1).", "class_sentence": "[1, 1, 2, 1, 1]", "sentences": "['Table 1 reports the results of the approaches we described in Section 4.', 'Out of the Kneser-Ney n-gram models, we found that the FST-C-9GRAMKN and the version modelling word boundaries (FST-Cwb-9GRAM-KN) to perform best on the synthetic parallel corpus and newspapers, respectively.', 'Cwb was not applied to the synthetic parallel corpus as we did not model word splitting.', 'The hybrid models (FST-RNNLM) outperformed all FST-only approaches.', 'Thus, we found the wordsplitting hybrid model (FST-RNNLM-Cwb) to be the best performing model overall (Table 1).']", "header_mention": "[None, ['FST-(C/Cwb)-9GRAM-KN', 'Newspaper 1', 'Newspaper 2', 'Corpus'], None, ['FST-RNNLM-(C/Cwb)'], ['FST-RNNLM-(C/Cwb)']]", "valid": 1, "image_id": "D18-1533-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance (%CERR). Slash-separated pairs denote FSTs incapable/capable of inserting word boundaries, respectively; see Section 4. The -KN suffix denotes Kneser-Ney smoothing. The data from Section 3.2 is used for evaluating the modern-orthography language model perplexity, and \u201cCorpus\u201d evaluates test-set transliteration performance from the synthetic missionary text back to the original modern text.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n LM perplexity || Valid.LM perplexity || TestTransliteration performance (%CERR) || CorpusTransliteration performance (%CERR) || Newspaper 1Transliteration performance (%CERR) || Newspaper 2
Transliteration model || FST-(C/Cwb)-7GRAM-KN3.073.1327.3%50.1% / 38.7%52.0% / 47.5%
Transliteration model || FST-(C/Cwb)-9GRAM-KN2.953.0226.6%50.7% / 39.3%52.5% / 47.2%
Transliteration model || FST-(C/Cwb)-11GRAM-KN2.943.0227.8%53.9% / 41.3%54.1% / 48.7%
Transliteration model || FST-RNNLM-(C/Cwb)2.652.6916.3%47.2% / 34.3%49.8% / 41.2%
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance (\\%CERR). Slash-separated pairs denote FSTs incapable/capable of inserting word boundaries, respectively; see Section 4. The -KN suffix denotes Kneser-Ney smoothing. The data from Section 3.2 is used for evaluating the modern-orthography language model perplexity, and \u201cCorpus\u201d evaluates test-set transliteration performance from the synthetic missionary text back to the original modern text.}\n\\begin{tabular}{lrrlll}\n\\toprule\n & LM perplexity $\\parallel$ Valid. & LM perplexity $\\parallel$ Test & Transliteration performance (\\%CERR) $\\parallel$ Corpus & Transliteration performance (\\%CERR) $\\parallel$ Newspaper 1 & Transliteration performance (\\%CERR) $\\parallel$ Newspaper 2 \\\\\n\\midrule\nTransliteration model $\\parallel$ FST-(C/Cwb)-7GRAM-KN & 3.07 & 3.13 & 27.3\\% & 50.1\\% / 38.7\\% & 52.0\\% / 47.5\\% \\\\\nTransliteration model $\\parallel$ FST-(C/Cwb)-9GRAM-KN & 2.95 & 3.02 & 26.6\\% & 50.7\\% / 39.3\\% & 52.5\\% / 47.2\\% \\\\\nTransliteration model $\\parallel$ FST-(C/Cwb)-11GRAM-KN & 2.94 & 3.02 & 27.8\\% & 53.9\\% / 41.3\\% & 54.1\\% / 48.7\\% \\\\\nTransliteration model $\\parallel$ FST-RNNLM-(C/Cwb) & 2.65 & 2.69 & 16.3\\% & 47.2\\% / 34.3\\% & 49.8\\% / 41.2\\% \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Performance (%CERR). Slash-separated pairs denote FSTs incapable/capable of inserting word boundaries, respectively; see Section 4. The -KN suffix denotes Kneser-Ney smoothing. The data from Section 3.2 is used for evaluating the modern-orthography language model perplexity, and \u201cCorpus\u201d evaluates test-set transliteration performance from the synthetic missionary text back to the original modern text.\n
\n \n LM perplexity || Valid.\n \n LM perplexity || Test\n \n Transliteration performance (%CERR) || Corpus\n \n Transliteration performance (%CERR) || Newspaper 1\n \n Transliteration performance (%CERR) || Newspaper 2\n
\n Transliteration model || FST-(C/Cwb)-7GRAM-KN\n \n 3.07\n \n 3.13\n \n 27.3%\n \n 50.1% / 38.7%\n \n 52.0% / 47.5%\n
\n Transliteration model || FST-(C/Cwb)-9GRAM-KN\n \n 2.95\n \n 3.02\n \n 26.6%\n \n 50.7% / 39.3%\n \n 52.5% / 47.2%\n
\n Transliteration model || FST-(C/Cwb)-11GRAM-KN\n \n 2.94\n \n 3.02\n \n 27.8%\n \n 53.9% / 41.3%\n \n 54.1% / 48.7%\n
\n Transliteration model || FST-RNNLM-(C/Cwb)\n \n 2.65\n \n 2.69\n \n 16.3%\n \n 47.2% / 34.3%\n \n 49.8% / 41.2%\n
\n" }, { "instance_id": "fb9152bc-e618-40ce-add4-720a185b9ffa", "table_id_paper": "D19-1037table_4", "caption": "Comparison of AUC Results", "row_header_level": 2, "row_headers": [ [ "Model", "PCNN+ATT" ], [ "Model", "PCNN+ONE" ] ], "column_header_level": 1, "column_headers": [ [ "-" ], [ "+SelfAtt" ], [ "CCL-CT" ] ], "contents": [ [ "0.341", "0.368", "0.381" ], [ "0.325", "0.352", "0.38" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy']", "target_entity": "['+SelfAtt', 'CCL-CT']", "table_html": "
- +SelfAtt CCL-CT
Model || PCNN+ATT 0.341 0.368 0.381
Model || PCNN+ONE 0.325 0.352 0.38
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D19-1037", "page_no": 8, "dir": "emnlp2019", "description": "Table 4 gives another comparison using AUC with all p-values less than 5e-02 from t-test evaluation. The results indicate that the larger AUC, the better performance. A simple ensemble model of two networks (AUC: 0.371) has a similar result as a single model (NetAtt, AUC: 0.368). The main purpose of adding an additional student network is to introduce con\u00ef\u00ac\u201aicts and to build the collaborative curriculum learning scheme. With the CCL strategies, our models improve performances by removing 'hard' (noisy) entity pair bags during.", "class_sentence": "[1, 1, 1, 2, 1]", "sentences": "['Table 4 gives another comparison using AUC with all p-values less than 5e-02 from t-test evaluation.', 'The results indicate that the larger AUC, the better performance.', 'A simple ensemble model of two networks (AUC: 0.371) has a similar result as a single model (NetAtt, AUC: 0.368).', 'The main purpose of adding an additional student network is to introduce con\u00ef\u00ac\u201aicts and to build the collaborative curriculum learning scheme.', \"With the CCL strategies, our models improve performances by removing 'hard' (noisy) entity pair bags during.\"]", "header_mention": "[None, None, ['+SelfAtt'], None, ['CCL-CT']]", "valid": 1, "image_id": "D19-1037-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison of AUC Results

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n -+SelfAttCCL-CT
Model || PCNN+ATT0.3410.3680.381
Model || PCNN+ONE0.3250.3520.38
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison of AUC Results}\n\\begin{tabular}{lrrr}\n\\toprule\n & - & +SelfAtt & CCL-CT \\\\\n\\midrule\nModel $\\parallel$ PCNN+ATT & 0.341 & 0.368 & 0.381 \\\\\nModel $\\parallel$ PCNN+ONE & 0.325 & 0.352 & 0.38 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Comparison of AUC Results\n
\n \n -\n \n +SelfAtt\n \n CCL-CT\n
\n Model || PCNN+ATT\n \n 0.341\n \n 0.368\n \n 0.381\n
\n Model || PCNN+ONE\n \n 0.325\n \n 0.352\n \n 0.38\n
\n" }, { "instance_id": "5723e795-b4e1-4f41-9f2f-d3e11dc9bd73", "table_id_paper": "D19-1068table_3", "caption": "Experimental results in exploring the shared syntactic order event detector.", "row_header_level": 2, "row_headers": [ [ "Model", "CL_Trans_MLP" ], [ "Model", "CL_Trans_CNN" ], [ "Model", "CL_Trans_Hbrid" ], [ "Model", "CL_Trans_Self." ], [ "Model", "CL_Trans_GCN (ours)" ], [ "Model", "CL_Trans_GCN_Self" ] ], "column_header_level": 1, "column_headers": [ [ "Pre." ], [ "Rec." ], [ "F1" ] ], "contents": [ [ "20.3", "16.3", "18.1" ], [ "32.5", "16.3", "21.7" ], [ "30.4", "17.6", "22.3" ], [ "34.9", "18.3", "24" ], [ "32", "23.4", "27" ], [ "32.1", "23", "26.8" ] ], "metrics_loc": "column", "metrics_type": "['Pre.', 'Rec.', 'F1']", "target_entity": "['CL_Trans_GCN (ours)']", "table_html": "
Pre. Rec. F1
Model || CL_Trans_MLP 20.3 16.3 18.1
Model || CL_Trans_CNN 32.5 16.3 21.7
Model || CL_Trans_Hbrid 30.4 17.6 22.3
Model || CL_Trans_Self. 34.9 18.3 24
Model || CL_Trans_GCN (ours) 32 23.4 27
Model || CL_Trans_GCN_Self 32.1 23 26.8
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D19-1068", "page_no": 8, "dir": "emnlp2019", "description": "Exploring the Syntactic Order Event Detector. We compare our syntactic order event detector (CL Trans GCN) with several event detectors, including 1) CL Trans MLP, which employs a feed-forward network as event detector; 2) CL Trans CNN, which uses CNNs as the event detector; 3) CL Trans Hbrid, which use a hybrid network (Feng et al., 2016) for event detection. We also compared our model with several variants including 4) CL Trans Self., which replaces the GCNs with a self-attention network, and 5) CL Trans GCN Self, which combines GCNs with a self-attention network. We train these models on the same translated English data. Table 3 shows the results. From the results, CL Trans MLP, CL Trans CNN, and CL Trans Hbrid behave poorly, as expected. The reason might be that these models usually employ order-sensitive structures (e.g., CNNs) for ED, which would suffer the word order inconsistency problem when trained on the translated data. CL Trans Self. yields relatively good performance. The reason might be that self-attention network could provide each word with a feature vector based on all the words of a sentence, which is also irrespective of the words\u00e2\u20ac\u2122 positions in a sentence. This could address the word order difference to some extent. Our syntactic order event detector yields the best performance.", "class_sentence": "[2, 2, 2, 2, 1, 1, 2, 1, 2, 2, 1]", "sentences": "['Exploring the Syntactic Order Event Detector.', 'We compare our syntactic order event detector (CL Trans GCN) with several event detectors, including 1) CL Trans MLP, which employs a feed-forward network as event detector; 2) CL Trans CNN, which uses CNNs as the event detector; 3) CL Trans Hbrid, which use a hybrid network (Feng et al., 2016) for event detection.', 'We also compared our model with several variants including 4) CL Trans Self., which replaces the GCNs with a self-attention network, and 5) CL Trans GCN Self, which combines GCNs with a self-attention network.', 'We train these models on the same translated English data.', 'Table 3 shows the results.', 'From the results, CL Trans MLP, CL Trans CNN, and CL Trans Hbrid behave poorly, as expected.', 'The reason might be that these models usually employ order-sensitive structures (e.g., CNNs) for ED, which would suffer the word order inconsistency problem when trained on the translated data.', 'CL Trans Self. yields relatively good performance.', 'The reason might be that self-attention network could provide each word with a feature vector based on all the words of a sentence, which is also irrespective of the words\u00e2\u20ac\u2122 positions in a sentence.', 'This could address the word order difference to some extent.', 'Our syntactic order event detector yields the best performance.']", "header_mention": "[None, None, None, None, None, ['CL_Trans_MLP', 'CL_Trans_CNN', 'CL_Trans_Hbrid'], None, ['CL_Trans_Self.'], None, None, ['CL_Trans_GCN (ours)']]", "valid": 1, "image_id": "D19-1068-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Experimental results in exploring the shared syntactic order event detector.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Pre.Rec.F1
Model || CL_Trans_MLP20.316.318.1
Model || CL_Trans_CNN32.516.321.7
Model || CL_Trans_Hbrid30.417.622.3
Model || CL_Trans_Self.34.918.324
Model || CL_Trans_GCN (ours)3223.427
Model || CL_Trans_GCN_Self32.12326.8
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Experimental results in exploring the shared syntactic order event detector.}\n\\begin{tabular}{lrrr}\n\\toprule\n & Pre. & Rec. & F1 \\\\\n\\midrule\nModel $\\parallel$ CL\\_Trans\\_MLP & 20.3 & 16.3 & 18.1 \\\\\nModel $\\parallel$ CL\\_Trans\\_CNN & 32.5 & 16.3 & 21.7 \\\\\nModel $\\parallel$ CL\\_Trans\\_Hbrid & 30.4 & 17.6 & 22.3 \\\\\nModel $\\parallel$ CL\\_Trans\\_Self. & 34.9 & 18.3 & 24 \\\\\nModel $\\parallel$ CL\\_Trans\\_GCN (ours) & 32 & 23.4 & 27 \\\\\nModel $\\parallel$ CL\\_Trans\\_GCN\\_Self & 32.1 & 23 & 26.8 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Experimental results in exploring the shared syntactic order event detector.\n
\n \n Pre.\n \n Rec.\n \n F1\n
\n Model || CL_Trans_MLP\n \n 20.3\n \n 16.3\n \n 18.1\n
\n Model || CL_Trans_CNN\n \n 32.5\n \n 16.3\n \n 21.7\n
\n Model || CL_Trans_Hbrid\n \n 30.4\n \n 17.6\n \n 22.3\n
\n Model || CL_Trans_Self.\n \n 34.9\n \n 18.3\n \n 24\n
\n Model || CL_Trans_GCN (ours)\n \n 32\n \n 23.4\n \n 27\n
\n Model || CL_Trans_GCN_Self\n \n 32.1\n \n 23\n \n 26.8\n
\n" }, { "instance_id": "f2a4db34-b17d-42be-bb29-d39aa81f6c03", "table_id_paper": "D19-1070table_8", "caption": "Results for each state change type. Performance on predicting creation and destruction are highest, partially due to the model\u2019s ability to use verb semantics for these tasks.", "row_header_level": 2, "row_headers": [ [ "Model", "ETBERT" ], [ "Model", "ETGPT" ] ], "column_header_level": 2, "column_headers": [ [ "Cat-1", "C" ], [ "Cat-2", "M" ], [ "Cat-3", "D" ], [ "Cat-2", "C" ], [ "Cat-3", "M" ], [ "Cat-4", "D" ] ], "contents": [ [ "78.51", "61.6", "71.5", "76.68", "54.12", "58.62" ], [ "79.82", "56.27", "73.83", "77.24", "50.82", "56.27" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy', 'accuracy']", "target_entity": "['C', 'M', 'D']", "table_html": "
Cat-1 || C Cat-2 || M Cat-3 || D Cat-2 || C Cat-3 || M Cat-4 || D
Model || ETBERT 78.51 61.6 71.5 76.68 54.12 58.62
Model || ETGPT 79.82 56.27 73.83 77.24 50.82 56.27
", "table_name": "Table 8", "table_id": "table_8", "paper_id": "D19-1070", "page_no": 8, "dir": "emnlp2019", "description": "For PROPARA, Table 5 shows that the model does not significantly outperform the SOTA models in state change detection (Cat-1). However, for those correctly detected events, the transformer model outperforms the previous models for detecting the exact step of state change (Cat-2), primarily based on verb semantics. We do a finer-grained study in Table 8 by breaking down the performance for the three state changes: creation (C), movement (M), and destruction (D), separately. Across the three state changes, the model suffers a loss of performance in the movement cases. This is owing to the fact that the movement cases require a deeper compositional and implicit event tracking. Also, a majority of errors leading to false negatives are due to the the formation of new sub-entities which are then mentioned with other names. For example, when talking about weak acid in \"he water becomes a weak acid. the water dissolves limestone\"\u009d the weak acid is also considered to move to the limestone.", "class_sentence": "[0, 0, 1, 1, 2, 2, 2]", "sentences": "['For PROPARA, Table 5 shows that the model does not significantly outperform the SOTA models in state change detection (Cat-1).', 'However, for those correctly detected events, the transformer model outperforms the previous models for detecting the exact step of state change (Cat-2), primarily based on verb semantics.', 'We do a finer-grained study in Table 8 by breaking down the performance for the three state changes: creation (C), movement (M), and destruction (D), separately.', 'Across the three state changes, the model suffers a loss of performance in the movement cases.', 'This is owing to the fact that the movement cases require a deeper compositional and implicit event tracking.', 'Also, a majority of errors leading to false negatives are due to the the formation of new sub-entities which are then mentioned with other names.', 'For example, when talking about weak acid in \"he water becomes a weak acid. the water dissolves limestone\"\\x9d the weak acid is also considered to move to the limestone.']", "header_mention": "[None, None, ['C', 'M', 'D'], ['M'], ['M'], None, None]", "valid": 1, "image_id": "D19-1070-Table8-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results for each state change type. Performance on predicting creation and destruction are highest, partially due to the model\u2019s ability to use verb semantics for these tasks.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Cat-1 || CCat-2 || MCat-3 || DCat-2 || CCat-3 || MCat-4 || D
Model || ETBERT78.5161.671.576.6854.1258.62
Model || ETGPT79.8256.2773.8377.2450.8256.27
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{7}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results for each state change type. Performance on predicting creation and destruction are highest, partially due to the model\u2019s ability to use verb semantics for these tasks.}\n\\begin{tabular}{lrrrrrr}\n\\toprule\n & Cat-1 $\\parallel$ C & Cat-2 $\\parallel$ M & Cat-3 $\\parallel$ D & Cat-2 $\\parallel$ C & Cat-3 $\\parallel$ M & Cat-4 $\\parallel$ D \\\\\n\\midrule\nModel $\\parallel$ ETBERT & 78.51 & 61.6 & 71.5 & 76.68 & 54.12 & 58.62 \\\\\nModel $\\parallel$ ETGPT & 79.82 & 56.27 & 73.83 & 77.24 & 50.82 & 56.27 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 8: Results for each state change type. Performance on predicting creation and destruction are highest, partially due to the model\u2019s ability to use verb semantics for these tasks.\n
\n \n Cat-1 || C\n \n Cat-2 || M\n \n Cat-3 || D\n \n Cat-2 || C\n \n Cat-3 || M\n \n Cat-4 || D\n
\n Model || ETBERT\n \n 78.51\n \n 61.6\n \n 71.5\n \n 76.68\n \n 54.12\n \n 58.62\n
\n Model || ETGPT\n \n 79.82\n \n 56.27\n \n 73.83\n \n 77.24\n \n 50.82\n \n 56.27\n
\n" }, { "instance_id": "05a20cea-21ae-4e86-a183-520be5a25428", "table_id_paper": "D19-1154table_1", "caption": "Comparison of multilingual sentence-image retrieval/matching (German-Image) and (English-Image) results on Multi30K. (Visual encoders:VGG\u2020 otherwise ResNet or Faster-RCNN(ResNet).) (Monolingual models*.)", "row_header_level": 2, "row_headers": [ [ "Method", "VSE\u2020* (Kiros et al., 2014)" ], [ "Method", "OE\u2020* (Vendrov et al., 2015)" ], [ "Method", "DAN* (Nam et al., 2017)" ], [ "Method", "VSE++* (Faghri et al., 2018)" ], [ "Method", "SCAN* (Lee et al., 2018)" ], [ "Method", "Pivot\u2020 (Gella et al., 2017)" ], [ "Method", "Ours\u2020 (Random, VGG19)" ], [ "Method", "Ours (Random, No diversity)" ], [ "Method", "Ours (Random)" ], [ "Method", "Ours (w/ FastText)" ], [ "Method", "Ours (w/ BERT)" ] ], "column_header_level": 2, "column_headers": [ [ "German to Image", "R@1" ], [ "German to Image", "R@5" ], [ "German to Image", "R10" ], [ "Image to German", "R@1" ], [ "Image to German", "R@5" ], [ "Image to German", "R10" ], [ "English to Image", "R@1" ], [ "English to Image", "R@5" ], [ "English to Image", "R10" ], [ "Image to English", "R@1" ], [ "Image to English", "R@5" ], [ "Image to English", "R10" ] ], "contents": [ [ "20.3", "47.2", "60.1", "29.3", "58.1", "71.8", "23.3", "53.6", "65.8", "31.6", "60.4", "72.7" ], [ "21.0", "48.5", "60.4", "26.8", "57.5", "70.9", "25.8", "56.5", "67.8", "34.8", "63.7", "74.8" ], [ "31.0", "60.9", "71.0", "46.5", "77.5", "83.0", "39.4", "69.2", "69.1", "55.0", "81.8", "89.0" ], [ "31.3", "62.2", "70.9", "47.5", "78.5", "84.5", "39.6", "69.1", "79.8", "53.1", "82.1", "87.5" ], [ "35.7", "64.9", "74.6", "52.3", "81.8", "88.5", "45.8", "74.4", "83.0", "61.8", "87.5", "93.7" ], [ "22.5", "49.3", "61.7", "28.2", "61.9", "73.4", "26.2", "56.4", "68.4", "33.8", "62.8", "75.2" ], [ "25.8", "54.9", "65.1", "34.1", "65.5", "76.5", "30.1", "62.5", "71.6", "36.4", "68.0", "80.9" ], [ "36.3", "65.3", "74.7", "53.1", "82.3", "88.8", "46.2", "74.7", "82.9", "63.3", "87.0", "93.3" ], [ "39.2", "67.5", "76.7", "55.0", "84.7", "91.2", "48.7", "77.2", "85.0", "66.4", "88.3", "93.4" ], [ "40.3", "70.1", "79.0", "60.4", "85.4", "92.0", "50.1", "78.1", "85.7", "68.0", "88.8", "94.0" ], [ "40.7", "70.5", "78.8", "56.5", "84.6", "91.3", "48.9", "78.3", "85.8", "66.5", "89.1", "94.1" ] ], "metrics_loc": "column", "metrics_type": "['R@1', 'R@5', 'R10', 'R@1', 'R@5', 'R10', 'R@1', 'R@5', 'R10', 'R@1', 'R@5', 'R10']", "target_entity": "['Ours (w/ FastText)', 'Ours (w/ BERT)']", "table_html": "
German to Image || R@1 German to Image || R@5 German to Image || R10 Image to German || R@1 Image to German || R@5 Image to German || R10 English to Image || R@1 English to Image || R@5 English to Image || R10 Image to English || R@1 Image to English || R@5 Image to English || R10
Method || VSE\u2020* (Kiros et al., 2014) 20.3 47.2 60.1 29.3 58.1 71.8 23.3 53.6 65.8 31.6 60.4 72.7
Method || OE\u2020* (Vendrov et al., 2015) 21.0 48.5 60.4 26.8 57.5 70.9 25.8 56.5 67.8 34.8 63.7 74.8
Method || DAN* (Nam et al., 2017) 31.0 60.9 71.0 46.5 77.5 83.0 39.4 69.2 69.1 55.0 81.8 89.0
Method || VSE++* (Faghri et al., 2018) 31.3 62.2 70.9 47.5 78.5 84.5 39.6 69.1 79.8 53.1 82.1 87.5
Method || SCAN* (Lee et al., 2018) 35.7 64.9 74.6 52.3 81.8 88.5 45.8 74.4 83.0 61.8 87.5 93.7
Method || Pivot\u2020 (Gella et al., 2017) 22.5 49.3 61.7 28.2 61.9 73.4 26.2 56.4 68.4 33.8 62.8 75.2
Method || Ours\u2020 (Random, VGG19) 25.8 54.9 65.1 34.1 65.5 76.5 30.1 62.5 71.6 36.4 68.0 80.9
Method || Ours (Random, No diversity) 36.3 65.3 74.7 53.1 82.3 88.8 46.2 74.7 82.9 63.3 87.0 93.3
Method || Ours (Random) 39.2 67.5 76.7 55.0 84.7 91.2 48.7 77.2 85.0 66.4 88.3 93.4
Method || Ours (w/ FastText) 40.3 70.1 79.0 60.4 85.4 92.0 50.1 78.1 85.7 68.0 88.8 94.0
Method || Ours (w/ BERT) 40.7 70.5 78.8 56.5 84.6 91.3 48.9 78.3 85.8 66.5 89.1 94.1
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "D19-1154", "page_no": 4, "dir": "emnlp2019", "description": "Table 1 presents the results on the Multi30K testing set. The VSE baselines in the first five rows are trained with English and German descriptions independently. In contrast, PIVOT (Gella et al., 2017) and the proposed model are capable of handling multilingual input queries with single model. For a fair comparison with PIVOT, we also report the result of swapping Faster-RCNN with VGG as the visual feature encoder in our model. As can be seen, the proposed models successfully obtain state-of-the-art results, outperforming other baselines by a significant margin. GermanImage matching benefit more from joint training with English-Image pairs. The models with pretrained multilingual embeddings and contextualized embeddings achieve better performance in comparison to randomly initialized word embeddings, especially for German. One explanation is that the degradation from German singletons is alleviated by the multi-task training with English and the pre-trained embeddings. While the model with BERT performs better in English, FastText is preferred for German-Image matching.", "class_sentence": "[1, 2, 1, 1, 1, 2, 2, 2, 1]", "sentences": "['Table 1 presents the results on the Multi30K testing set.', 'The VSE baselines in the first five rows are trained with English and German descriptions independently.', 'In contrast, PIVOT (Gella et al., 2017) and the proposed model are capable of handling multilingual input queries with single model.', 'For a fair comparison with PIVOT, we also report the result of swapping Faster-RCNN with VGG as the visual feature encoder in our model.', 'As can be seen, the proposed models successfully obtain state-of-the-art results, outperforming other baselines by a significant margin.', 'GermanImage matching benefit more from joint training with English-Image pairs.', 'The models with pretrained multilingual embeddings and contextualized embeddings achieve better performance in comparison to randomly initialized word embeddings, especially for German.', 'One explanation is that the degradation from German singletons is alleviated by the multi-task training with English and the pre-trained embeddings.', 'While the model with BERT performs better in English, FastText is preferred for German-Image matching.']", "header_mention": "[None, ['VSE\u2020* (Kiros et al., 2014)', 'VSE++* (Faghri et al., 2018)'], ['Pivot\u2020 (Gella et al., 2017)', 'Ours\u2020 (Random, VGG19)', 'Ours (Random, No diversity)', 'Ours (Random)', 'Ours (w/ FastText)', 'Ours (w/ BERT)'], ['Pivot\u2020 (Gella et al., 2017)'], ['Ours\u2020 (Random, VGG19)', 'Ours (Random, No diversity)', 'Ours (Random)', 'Ours (w/ FastText)', 'Ours (w/ BERT)'], None, None, None, ['Ours (w/ FastText)', 'Ours (w/ BERT)', 'English to Image', 'Image to English', 'German to Image', 'Image to German']]", "valid": 1, "image_id": "D19-1154-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison of multilingual sentence-image retrieval/matching (German-Image) and (English-Image) results on Multi30K. (Visual encoders:VGG\u2020 otherwise ResNet or Faster-RCNN(ResNet).) (Monolingual models*.)

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n German to Image || R@1German to Image || R@5German to Image || R10Image to German || R@1Image to German || R@5Image to German || R10English to Image || R@1English to Image || R@5English to Image || R10Image to English || R@1Image to English || R@5Image to English || R10
Method || VSE\u2020* (Kiros et al., 2014)20.347.260.129.358.171.823.353.665.831.660.472.7
Method || OE\u2020* (Vendrov et al., 2015)21.048.560.426.857.570.925.856.567.834.863.774.8
Method || DAN* (Nam et al., 2017)31.060.971.046.577.583.039.469.269.155.081.889.0
Method || VSE++* (Faghri et al., 2018)31.362.270.947.578.584.539.669.179.853.182.187.5
Method || SCAN* (Lee et al., 2018)35.764.974.652.381.888.545.874.483.061.887.593.7
Method || Pivot\u2020 (Gella et al., 2017)22.549.361.728.261.973.426.256.468.433.862.875.2
Method || Ours\u2020 (Random, VGG19)25.854.965.134.165.576.530.162.571.636.468.080.9
Method || Ours (Random, No diversity)36.365.374.753.182.388.846.274.782.963.387.093.3
Method || Ours (Random)39.267.576.755.084.791.248.777.285.066.488.393.4
Method || Ours (w/ FastText)40.370.179.060.485.492.050.178.185.768.088.894.0
Method || Ours (w/ BERT)40.770.578.856.584.691.348.978.385.866.589.194.1
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison of multilingual sentence-image retrieval/matching (German-Image) and (English-Image) results on Multi30K. (Visual encoders:VGG$\\dagger$ otherwise ResNet or Faster-RCNN(ResNet).) (Monolingual models*.)}\n\\begin{tabular}{lrrrrrrrrrrrr}\n\\toprule\n & German to Image $\\parallel$ R@1 & German to Image $\\parallel$ R@5 & German to Image $\\parallel$ R10 & Image to German $\\parallel$ R@1 & Image to German $\\parallel$ R@5 & Image to German $\\parallel$ R10 & English to Image $\\parallel$ R@1 & English to Image $\\parallel$ R@5 & English to Image $\\parallel$ R10 & Image to English $\\parallel$ R@1 & Image to English $\\parallel$ R@5 & Image to English $\\parallel$ R10 \\\\\n\\midrule\nMethod $\\parallel$ VSE$\\dagger$* (Kiros et al., 2014) & 20.3 & 47.2 & 60.1 & 29.3 & 58.1 & 71.8 & 23.3 & 53.6 & 65.8 & 31.6 & 60.4 & 72.7 \\\\\nMethod $\\parallel$ OE$\\dagger$* (Vendrov et al., 2015) & 21 & 48.5 & 60.4 & 26.8 & 57.5 & 70.9 & 25.8 & 56.5 & 67.8 & 34.8 & 63.7 & 74.8 \\\\\nMethod $\\parallel$ DAN* (Nam et al., 2017) & 31 & 60.9 & 71 & 46.5 & 77.5 & 83 & 39.4 & 69.2 & 69.1 & 55 & 81.8 & 89 \\\\\nMethod $\\parallel$ VSE++* (Faghri et al., 2018) & 31.3 & 62.2 & 70.9 & 47.5 & 78.5 & 84.5 & 39.6 & 69.1 & 79.8 & 53.1 & 82.1 & 87.5 \\\\\nMethod $\\parallel$ SCAN* (Lee et al., 2018) & 35.7 & 64.9 & 74.6 & 52.3 & 81.8 & 88.5 & 45.8 & 74.4 & 83 & 61.8 & 87.5 & 93.7 \\\\\nMethod $\\parallel$ Pivot$\\dagger$ (Gella et al., 2017) & 22.5 & 49.3 & 61.7 & 28.2 & 61.9 & 73.4 & 26.2 & 56.4 & 68.4 & 33.8 & 62.8 & 75.2 \\\\\nMethod $\\parallel$ Ours$\\dagger$ (Random, VGG19) & 25.8 & 54.9 & 65.1 & 34.1 & 65.5 & 76.5 & 30.1 & 62.5 & 71.6 & 36.4 & 68 & 80.9 \\\\\nMethod $\\parallel$ Ours (Random, No diversity) & 36.3 & 65.3 & 74.7 & 53.1 & 82.3 & 88.8 & 46.2 & 74.7 & 82.9 & 63.3 & 87 & 93.3 \\\\\nMethod $\\parallel$ Ours (Random) & 39.2 & 67.5 & 76.7 & 55 & 84.7 & 91.2 & 48.7 & 77.2 & 85 & 66.4 & 88.3 & 93.4 \\\\\nMethod $\\parallel$ Ours (w/ FastText) & 40.3 & 70.1 & 79 & 60.4 & 85.4 & 92 & 50.1 & 78.1 & 85.7 & 68 & 88.8 & 94 \\\\\nMethod $\\parallel$ Ours (w/ BERT) & 40.7 & 70.5 & 78.8 & 56.5 & 84.6 & 91.3 & 48.9 & 78.3 & 85.8 & 66.5 & 89.1 & 94.1 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Comparison of multilingual sentence-image retrieval/matching (German-Image) and (English-Image) results on Multi30K. (Visual encoders:VGG\u2020 otherwise ResNet or Faster-RCNN(ResNet).) (Monolingual models*.)\n
\n \n German to Image || R@1\n \n German to Image || R@5\n \n German to Image || R10\n \n Image to German || R@1\n \n Image to German || R@5\n \n Image to German || R10\n \n English to Image || R@1\n \n English to Image || R@5\n \n English to Image || R10\n \n Image to English || R@1\n \n Image to English || R@5\n \n Image to English || R10\n
\n Method || VSE\u2020* (Kiros et al., 2014)\n \n 20.3\n \n 47.2\n \n 60.1\n \n 29.3\n \n 58.1\n \n 71.8\n \n 23.3\n \n 53.6\n \n 65.8\n \n 31.6\n \n 60.4\n \n 72.7\n
\n Method || OE\u2020* (Vendrov et al., 2015)\n \n 21.0\n \n 48.5\n \n 60.4\n \n 26.8\n \n 57.5\n \n 70.9\n \n 25.8\n \n 56.5\n \n 67.8\n \n 34.8\n \n 63.7\n \n 74.8\n
\n Method || DAN* (Nam et al., 2017)\n \n 31.0\n \n 60.9\n \n 71.0\n \n 46.5\n \n 77.5\n \n 83.0\n \n 39.4\n \n 69.2\n \n 69.1\n \n 55.0\n \n 81.8\n \n 89.0\n
\n Method || VSE++* (Faghri et al., 2018)\n \n 31.3\n \n 62.2\n \n 70.9\n \n 47.5\n \n 78.5\n \n 84.5\n \n 39.6\n \n 69.1\n \n 79.8\n \n 53.1\n \n 82.1\n \n 87.5\n
\n Method || SCAN* (Lee et al., 2018)\n \n 35.7\n \n 64.9\n \n 74.6\n \n 52.3\n \n 81.8\n \n 88.5\n \n 45.8\n \n 74.4\n \n 83.0\n \n 61.8\n \n 87.5\n \n 93.7\n
\n Method || Pivot\u2020 (Gella et al., 2017)\n \n 22.5\n \n 49.3\n \n 61.7\n \n 28.2\n \n 61.9\n \n 73.4\n \n 26.2\n \n 56.4\n \n 68.4\n \n 33.8\n \n 62.8\n \n 75.2\n
\n Method || Ours\u2020 (Random, VGG19)\n \n 25.8\n \n 54.9\n \n 65.1\n \n 34.1\n \n 65.5\n \n 76.5\n \n 30.1\n \n 62.5\n \n 71.6\n \n 36.4\n \n 68.0\n \n 80.9\n
\n Method || Ours (Random, No diversity)\n \n 36.3\n \n 65.3\n \n 74.7\n \n 53.1\n \n 82.3\n \n 88.8\n \n 46.2\n \n 74.7\n \n 82.9\n \n 63.3\n \n 87.0\n \n 93.3\n
\n Method || Ours (Random)\n \n 39.2\n \n 67.5\n \n 76.7\n \n 55.0\n \n 84.7\n \n 91.2\n \n 48.7\n \n 77.2\n \n 85.0\n \n 66.4\n \n 88.3\n \n 93.4\n
\n Method || Ours (w/ FastText)\n \n 40.3\n \n 70.1\n \n 79.0\n \n 60.4\n \n 85.4\n \n 92.0\n \n 50.1\n \n 78.1\n \n 85.7\n \n 68.0\n \n 88.8\n \n 94.0\n
\n Method || Ours (w/ BERT)\n \n 40.7\n \n 70.5\n \n 78.8\n \n 56.5\n \n 84.6\n \n 91.3\n \n 48.9\n \n 78.3\n \n 85.8\n \n 66.5\n \n 89.1\n \n 94.1\n
\n" }, { "instance_id": "789efed8-7113-45b4-b0eb-3a5e8c1bf234", "table_id_paper": "D19-1166table_8", "caption": "Evaluation on auxiliary tasks: Multitask models trained on both the translation and simplification dataset improves the performance for the task of English Simplification.", "row_header_level": 3, "row_headers": [ [ "Task", "English Simplification", "Simplify" ], [ "Task", "English Simplification", "Translate and Simplify" ], [ "Task", "English Simplification", "All Tasks" ], [ "Task", "Machine Translation", "Translate" ], [ "Task", "Machine Translation", "Translate and Simplify" ], [ "Task", "Machine Translation", "All Tasks" ] ], "column_header_level": 1, "column_headers": [ [ "BLEU" ], [ "SARI" ], [ "PCC" ] ], "contents": [ [ "55.76", "41.7", "0.736" ], [ "56.47", "41.3", "0.730" ], [ "56.05", "42.1", "0.736" ], [ "29.09", "-", "0.769" ], [ "27.33", "-", "0.647" ], [ "27.63", "-", "0.658" ] ], "metrics_loc": "column", "metrics_type": "['BLEU', 'SARI', 'PCC']", "target_entity": "['Machine Translation']", "table_html": "
BLEU SARI PCC
Task || English Simplification || Simplify 55.76 41.7 0.736
Task || English Simplification || Translate and Simplify 56.47 41.3 0.730
Task || English Simplification || All Tasks 56.05 42.1 0.736
Task || Machine Translation || Translate 29.09 - 0.769
Task || Machine Translation || Translate and Simplify 27.33 - 0.647
Task || Machine Translation || All Tasks 27.63 - 0.658
", "table_name": "Table 8", "table_id": "table_8", "paper_id": "D19-1166", "page_no": 9, "dir": "emnlp2019", "description": "Table 8 summarizes the results: the multitask model slightly outperforms a dedicated simplification model on English simplification, showing the benefits of the additional training data from other tasks. By contrast, on the resource-rich MT task, the standalone translation system performs better. This can be explained by the fact that the standalone system is only responsible for text translation, while the multi-task model is exposed to samples of more diverse complexity levels during training which damage its ability to preserve complexity.", "class_sentence": "[1, 1, 2]", "sentences": "['Table 8 summarizes the results: the multitask model slightly outperforms a dedicated simplification model on English simplification, showing the benefits of the additional training data from other tasks.', 'By contrast, on the resource-rich MT task, the standalone translation system performs better.', 'This can be explained by the fact that the standalone system is only responsible for text translation, while the multi-task model is exposed to samples of more diverse complexity levels during training which damage its ability to preserve complexity.']", "header_mention": "[['Machine Translation', 'English Simplification'], ['Machine Translation', 'Translate'], None]", "valid": 1, "image_id": "D19-1166-Table8-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Evaluation on auxiliary tasks: Multitask models trained on both the translation and simplification dataset improves the performance for the task of English Simplification.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n BLEUSARIPCC
Task || English Simplification || Simplify55.7641.70.736
Task || English Simplification || Translate and Simplify56.4741.30.730
Task || English Simplification || All Tasks56.0542.10.736
Task || Machine Translation || Translate29.09-0.769
Task || Machine Translation || Translate and Simplify27.33-0.647
Task || Machine Translation || All Tasks27.63-0.658
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{7}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Evaluation on auxiliary tasks: Multitask models trained on both the translation and simplification dataset improves the performance for the task of English Simplification.}\n\\begin{tabular}{lrlr}\n\\toprule\n & BLEU & SARI & PCC \\\\\n\\midrule\nTask $\\parallel$ English Simplification $\\parallel$ Simplify & 55.76 & 41.7 & 0.736 \\\\\nTask $\\parallel$ English Simplification $\\parallel$ Translate and Simplify & 56.47 & 41.3 & 0.73 \\\\\nTask $\\parallel$ English Simplification $\\parallel$ All Tasks & 56.05 & 42.1 & 0.736 \\\\\nTask $\\parallel$ Machine Translation $\\parallel$ Translate & 29.09 & - & 0.769 \\\\\nTask $\\parallel$ Machine Translation $\\parallel$ Translate and Simplify & 27.33 & - & 0.647 \\\\\nTask $\\parallel$ Machine Translation $\\parallel$ All Tasks & 27.63 & - & 0.658 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 8: Evaluation on auxiliary tasks: Multitask models trained on both the translation and simplification dataset improves the performance for the task of English Simplification.\n
\n \n BLEU\n \n SARI\n \n PCC\n
\n Task || English Simplification || Simplify\n \n 55.76\n \n 41.7\n \n 0.736\n
\n Task || English Simplification || Translate and Simplify\n \n 56.47\n \n 41.3\n \n 0.730\n
\n Task || English Simplification || All Tasks\n \n 56.05\n \n 42.1\n \n 0.736\n
\n Task || Machine Translation || Translate\n \n 29.09\n \n -\n \n 0.769\n
\n Task || Machine Translation || Translate and Simplify\n \n 27.33\n \n -\n \n 0.647\n
\n Task || Machine Translation || All Tasks\n \n 27.63\n \n -\n \n 0.658\n
\n" }, { "instance_id": "28adcf93-4a9b-4236-9eb5-e0c93234b0d2", "table_id_paper": "D19-1174table_2", "caption": "Results with traditional machine learning (Label Powerset)", "row_header_level": 2, "row_headers": [ [ "Classifier", "SVM" ], [ "Classifier", "LR" ], [ "Classifier", "RF" ] ], "column_header_level": 3, "column_headers": [ [ "Features", "Word n-grams", "FI" ], [ "Features", "Word n-grams", "Fmacro" ], [ "Features", "Word n-grams", "AccI" ], [ "Features", "Word n-grams", "Fmicro" ], [ "Features", "Character n-grams", "FI" ], [ "Features", "Character n-grams", "Fmacro" ], [ "Features", "Character n-grams", "AccI" ], [ "Features", "Character n-grams", "Fmicro" ], [ "Features", "Averaged ELMo vectors", "FI" ], [ "Features", "Averaged ELMo vectors", "Fmacro" ], [ "Features", "Averaged ELMo vectors", "AccI" ], [ "Features", "Averaged ELMo vectors", "Fmicro" ], [ "Features", "Composite features", "FI" ], [ "Features", "Composite features", "Fmacro" ], [ "Features", "Composite features", "AccI" ], [ "Features", "Composite features", "Fmicro" ] ], "contents": [ [ "0.448", "0.373", "0.324", "0.410", "0.449", "0.374", "0.331", "0.416", "0.546", "0.430", "0.431", "0.500", "0.178", "0.094", "0.116", "0.174" ], [ "0.357", "0.315", "0.236", "0.349", "0.357", "0.311", "0.230", "0.352", "0.595", "0.479", "0.478", "0.549", "0.438", "0.370", "0.311", "0.421" ], [ "0.531", "0.398", "0.438", "0.476", "0.395", "0.205", "0.325", "0.349", "0.375", "0.164", "0.305", "0.331", "0.460", "0.311", "0.380", "0.415" ] ], "metrics_loc": "column", "metrics_type": "['FI', 'Fmacro', 'AccI', 'Fmicro', 'FI', 'Fmacro', 'AccI', 'Fmicro', 'FI', 'Fmacro', 'AccI', 'Fmicro', 'FI', 'Fmacro', 'AccI', 'Fmicro']", "target_entity": "['Classifier']", "table_html": "
Features || Word n-grams || FI Features || Word n-grams || Fmacro Features || Word n-grams || AccI Features || Word n-grams || Fmicro Features || Character n-grams || FI Features || Character n-grams || Fmacro Features || Character n-grams || AccI Features || Character n-grams || Fmicro Features || Averaged ELMo vectors || FI Features || Averaged ELMo vectors || Fmacro Features || Averaged ELMo vectors || AccI Features || Averaged ELMo vectors || Fmicro Features || Composite features || FI Features || Composite features || Fmacro Features || Composite features || AccI Features || Composite features || Fmicro
Classifier || SVM 0.448 0.373 0.324 0.410 0.449 0.374 0.331 0.416 0.546 0.430 0.431 0.500 0.178 0.094 0.116 0.174
Classifier || LR 0.357 0.315 0.236 0.349 0.357 0.311 0.230 0.352 0.595 0.479 0.478 0.549 0.438 0.370 0.311 0.421
Classifier || RF 0.531 0.398 0.438 0.476 0.395 0.205 0.325 0.349 0.375 0.164 0.305 0.331 0.460 0.311 0.380 0.415
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D19-1174", "page_no": 8, "dir": "emnlp2019", "description": "5.3 Results. Table 2 shows results produced using traditional ML methods (SVM, RF, and LR) across four different feature sets (word n-grams, character ngrams, averaged ELMo vectors, and composite features). We use Label Powerset for these methods, since the direct (non-transformative) formulation cannot be used with them. Among these combinations, logistic regression with averaged ELMo embeddings as features performs the best.", "class_sentence": "[2, 1, 2, 1]", "sentences": "['5.3 Results.', 'Table 2 shows results produced using traditional ML methods (SVM, RF, and LR) across four different feature sets (word n-grams, character ngrams, averaged ELMo vectors, and composite features).', 'We use Label Powerset for these methods, since the direct (non-transformative) formulation cannot be used with them.', 'Among these combinations, logistic regression with averaged ELMo embeddings as features performs the best.']", "header_mention": "[None, ['SVM', 'RF', 'LR', 'Word n-grams', 'Averaged ELMo vectors', 'Composite features'], None, ['LR', 'Averaged ELMo vectors']]", "valid": 1, "image_id": "D19-1174-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results with traditional machine learning (Label Powerset)

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Features || Word n-grams || FIFeatures || Word n-grams || FmacroFeatures || Word n-grams || AccIFeatures || Word n-grams || FmicroFeatures || Character n-grams || FIFeatures || Character n-grams || FmacroFeatures || Character n-grams || AccIFeatures || Character n-grams || FmicroFeatures || Averaged ELMo vectors || FIFeatures || Averaged ELMo vectors || FmacroFeatures || Averaged ELMo vectors || AccIFeatures || Averaged ELMo vectors || FmicroFeatures || Composite features || FIFeatures || Composite features || FmacroFeatures || Composite features || AccIFeatures || Composite features || Fmicro
Classifier || SVM0.4480.3730.3240.4100.4490.3740.3310.4160.5460.4300.4310.5000.1780.0940.1160.174
Classifier || LR0.3570.3150.2360.3490.3570.3110.2300.3520.5950.4790.4780.5490.4380.3700.3110.421
Classifier || RF0.5310.3980.4380.4760.3950.2050.3250.3490.3750.1640.3050.3310.4600.3110.3800.415
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results with traditional machine learning (Label Powerset)}\n\\begin{tabular}{lrrrrrrrrrrrrrrrr}\n\\toprule\n & Features $\\parallel$ Word n-grams $\\parallel$ FI & Features $\\parallel$ Word n-grams $\\parallel$ Fmacro & Features $\\parallel$ Word n-grams $\\parallel$ AccI & Features $\\parallel$ Word n-grams $\\parallel$ Fmicro & Features $\\parallel$ Character n-grams $\\parallel$ FI & Features $\\parallel$ Character n-grams $\\parallel$ Fmacro & Features $\\parallel$ Character n-grams $\\parallel$ AccI & Features $\\parallel$ Character n-grams $\\parallel$ Fmicro & Features $\\parallel$ Averaged ELMo vectors $\\parallel$ FI & Features $\\parallel$ Averaged ELMo vectors $\\parallel$ Fmacro & Features $\\parallel$ Averaged ELMo vectors $\\parallel$ AccI & Features $\\parallel$ Averaged ELMo vectors $\\parallel$ Fmicro & Features $\\parallel$ Composite features $\\parallel$ FI & Features $\\parallel$ Composite features $\\parallel$ Fmacro & Features $\\parallel$ Composite features $\\parallel$ AccI & Features $\\parallel$ Composite features $\\parallel$ Fmicro \\\\\n\\midrule\nClassifier $\\parallel$ SVM & 0.448 & 0.373 & 0.324 & 0.41 & 0.449 & 0.374 & 0.331 & 0.416 & 0.546 & 0.43 & 0.431 & 0.5 & 0.178 & 0.094 & 0.116 & 0.174 \\\\\nClassifier $\\parallel$ LR & 0.357 & 0.315 & 0.236 & 0.349 & 0.357 & 0.311 & 0.23 & 0.352 & 0.595 & 0.479 & 0.478 & 0.549 & 0.438 & 0.37 & 0.311 & 0.421 \\\\\nClassifier $\\parallel$ RF & 0.531 & 0.398 & 0.438 & 0.476 & 0.395 & 0.205 & 0.325 & 0.349 & 0.375 & 0.164 & 0.305 & 0.331 & 0.46 & 0.311 & 0.38 & 0.415 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Results with traditional machine learning (Label Powerset)\n
\n \n Features || Word n-grams || FI\n \n Features || Word n-grams || Fmacro\n \n Features || Word n-grams || AccI\n \n Features || Word n-grams || Fmicro\n \n Features || Character n-grams || FI\n \n Features || Character n-grams || Fmacro\n \n Features || Character n-grams || AccI\n \n Features || Character n-grams || Fmicro\n \n Features || Averaged ELMo vectors || FI\n \n Features || Averaged ELMo vectors || Fmacro\n \n Features || Averaged ELMo vectors || AccI\n \n Features || Averaged ELMo vectors || Fmicro\n \n Features || Composite features || FI\n \n Features || Composite features || Fmacro\n \n Features || Composite features || AccI\n \n Features || Composite features || Fmicro\n
\n Classifier || SVM\n \n 0.448\n \n 0.373\n \n 0.324\n \n 0.410\n \n 0.449\n \n 0.374\n \n 0.331\n \n 0.416\n \n 0.546\n \n 0.430\n \n 0.431\n \n 0.500\n \n 0.178\n \n 0.094\n \n 0.116\n \n 0.174\n
\n Classifier || LR\n \n 0.357\n \n 0.315\n \n 0.236\n \n 0.349\n \n 0.357\n \n 0.311\n \n 0.230\n \n 0.352\n \n 0.595\n \n 0.479\n \n 0.478\n \n 0.549\n \n 0.438\n \n 0.370\n \n 0.311\n \n 0.421\n
\n Classifier || RF\n \n 0.531\n \n 0.398\n \n 0.438\n \n 0.476\n \n 0.395\n \n 0.205\n \n 0.325\n \n 0.349\n \n 0.375\n \n 0.164\n \n 0.305\n \n 0.331\n \n 0.460\n \n 0.311\n \n 0.380\n \n 0.415\n
\n" }, { "instance_id": "c4789ca0-bb05-4641-9679-8c15ca760b0a", "table_id_paper": "D19-1177table_3", "caption": "Results of performance (AUROC) of up-to-date methods on identifying helpful reviews evaluated by the test sets of Amazon. (italic f onts\u2217: the best performance among the baseline approaches; bold fonts: the state-of-the-art performance of all the approaches)", "row_header_level": 2, "row_headers": [ [ "Category (Amazon)", "Clothing, Shoes & Jewelry" ], [ "Category (Amazon)", "Electronics" ], [ "Category (Amazon)", "Grocery & Gourmet Food" ], [ "Category (Amazon)", "Health & Personal Care" ], [ "Category (Amazon)", "Home & Kitchen" ], [ "Category (Amazon)", "Movies & TV" ], [ "Category (Amazon)", "Pet Supplies" ], [ "Category (Amazon)", "Tools & Home Improvement" ], [ "Category (Amazon)", "MACRO AVERAGE" ], [ "Category (Amazon)", "MICRO AVERAGE (Primary)" ] ], "column_header_level": 2, "column_headers": [ [ "Area under Receiver Operating Characteristic (AUROC)", "STR" ], [ "Area under Receiver Operating Characteristic (AUROC)", "LEX" ], [ "Area under Receiver Operating Characteristic (AUROC)", "GALC" ], [ "Area under Receiver Operating Characteristic (AUROC)", "INQUIRER" ], [ "Area under Receiver Operating Characteristic (AUROC)", "FUSION (SVM)" ], [ "Area under Receiver Operating Characteristic (AUROC)", "FUSION (R.F.)" ], [ "Area under Receiver Operating Characteristic (AUROC)", "EG-CNN" ], [ "Area under Receiver Operating Characteristic (AUROC)", "MTNL" ], [ "Area under Receiver Operating Characteristic (AUROC)", "R2HP" ] ], "contents": [ [ "0.548", "0.536", "0.562", "0.601*", "0.579", "0.55", "0.583", "0.589", "0.623 (+0.022)" ], [ "0.583", "0.549", "0.588", "0.616*", "0.577", "0.58", "0.611", "0.613", "0.661 (+0.045)" ], [ "0.536", "0.526", "0.553", "0.602", "0.532", "0.546", "0.611", "0.626*", "0.657 (+0.031)" ], [ "0.558", "0.523", "0.559", "0.61", "0.591", "0.562", "0.613", "0.620*", "0.683 (+0.063)" ], [ "0.568", "0.537", "0.565", "0.597", "0.569", "0.573", "0.603", "0.610*", "0.646 (+0.036)" ], [ "0.603", "0.558", "0.621", "0.634", "0.603", "0.607", "0.648", "0.652*", "0.713 (+0.061)" ], [ "0.56", "0.542", "0.585", "0.603", "0.548", "0.558", "0.58", "0.629*", "0.692 (+0.063)" ], [ "0.584", "0.558", "0.58", "0.592", "0.575", "0.586", "0.617", "0.624*", "0.672 (+0.048)" ], [ "0.568", "0.541", "0.577", "0.607", "0.572", "0.57", "0.608", "0.620*", "0.668 (+0.048)" ], [ "0.571", "0.543", "0.58", "0.609", "0.573", "0.574", "0.614", "0.625*", "0.675 (+0.049)" ] ], "metrics_loc": "column", "metrics_type": "['Area under Receiver Operating Characteristic (AUROC)', 'Area under Receiver Operating Characteristic (AUROC)', 'Area under Receiver Operating Characteristic (AUROC)', 'Area under Receiver Operating Characteristic (AUROC)', 'Area under Receiver Operating Characteristic (AUROC)', 'Area under Receiver Operating Characteristic (AUROC)', 'Area under Receiver Operating Characteristic (AUROC)', 'Area under Receiver Operating Characteristic (AUROC)', 'Area under Receiver Operating Characteristic (AUROC)']", "target_entity": "['R2HP']", "table_html": "
Area under Receiver Operating Characteristic (AUROC) || STR Area under Receiver Operating Characteristic (AUROC) || LEX Area under Receiver Operating Characteristic (AUROC) || GALC Area under Receiver Operating Characteristic (AUROC) || INQUIRER Area under Receiver Operating Characteristic (AUROC) || FUSION (SVM) Area under Receiver Operating Characteristic (AUROC) || FUSION (R.F.) Area under Receiver Operating Characteristic (AUROC) || EG-CNN Area under Receiver Operating Characteristic (AUROC) || MTNL Area under Receiver Operating Characteristic (AUROC) || R2HP
Category (Amazon) || Clothing, Shoes & Jewelry 0.548 0.536 0.562 0.601* 0.579 0.55 0.583 0.589 0.623 (+0.022)
Category (Amazon) || Electronics 0.583 0.549 0.588 0.616* 0.577 0.58 0.611 0.613 0.661 (+0.045)
Category (Amazon) || Grocery & Gourmet Food 0.536 0.526 0.553 0.602 0.532 0.546 0.611 0.626* 0.657 (+0.031)
Category (Amazon) || Health & Personal Care 0.558 0.523 0.559 0.61 0.591 0.562 0.613 0.620* 0.683 (+0.063)
Category (Amazon) || Home & Kitchen 0.568 0.537 0.565 0.597 0.569 0.573 0.603 0.610* 0.646 (+0.036)
Category (Amazon) || Movies & TV 0.603 0.558 0.621 0.634 0.603 0.607 0.648 0.652* 0.713 (+0.061)
Category (Amazon) || Pet Supplies 0.56 0.542 0.585 0.603 0.548 0.558 0.58 0.629* 0.692 (+0.063)
Category (Amazon) || Tools & Home Improvement 0.584 0.558 0.58 0.592 0.575 0.586 0.617 0.624* 0.672 (+0.048)
Category (Amazon) || MACRO AVERAGE 0.568 0.541 0.577 0.607 0.572 0.57 0.608 0.620* 0.668 (+0.048)
Category (Amazon) || MICRO AVERAGE (Primary) 0.571 0.543 0.58 0.609 0.573 0.574 0.614 0.625* 0.675 (+0.049)
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D19-1177", "page_no": 8, "dir": "emnlp2019", "description": "Identification of helpful reviews: As both the training and test sets are imbalanced, we adopt the Area under Receiver Operating Characteristic (AUROC) as the metric to evaluate the performance of all the approaches on helpful review identification. In line with Table 3, MTNL (Fan et al., 2018) achieves the best performance up-to-date on this classification task among the baseline approaches as it achieves the best performance on 12 of 14 categories in Amazon and Yelp datasets. R2HP surpasses MTNL on both datasets and obtains state-of-the-art (microaveraged) results of 67.5% AUROC (Amazon) and 75.1% AUROC (Yelp) with absolute improvements of 4.9% AUROC and 4.7% AUROC, respectively.", "class_sentence": "[2, 1, 1]", "sentences": "['Identification of helpful reviews: As both the training and test sets are imbalanced, we adopt the Area under Receiver Operating Characteristic (AUROC) as the metric to evaluate the performance of all the approaches on helpful review identification.', 'In line with Table 3, MTNL (Fan et al., 2018) achieves the best performance up-to-date on this classification task among the baseline approaches as it achieves the best performance on 12 of 14 categories in Amazon and Yelp datasets.', 'R2HP surpasses MTNL on both datasets and obtains state-of-the-art (microaveraged) results of 67.5% AUROC (Amazon) and 75.1% AUROC (Yelp) with absolute improvements of 4.9% AUROC and 4.7% AUROC, respectively.']", "header_mention": "[None, ['MTNL'], ['R2HP', 'Area under Receiver Operating Characteristic (AUROC)']]", "valid": 1, "image_id": "D19-1177-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results of performance (AUROC) of up-to-date methods on identifying helpful reviews evaluated by the test sets of Amazon. (italic f onts\u2217: the best performance among the baseline approaches; bold fonts: the state-of-the-art performance of all the approaches)

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Area under Receiver Operating Characteristic (AUROC) || STRArea under Receiver Operating Characteristic (AUROC) || LEXArea under Receiver Operating Characteristic (AUROC) || GALCArea under Receiver Operating Characteristic (AUROC) || INQUIRERArea under Receiver Operating Characteristic (AUROC) || FUSION (SVM)Area under Receiver Operating Characteristic (AUROC) || FUSION (R.F.)Area under Receiver Operating Characteristic (AUROC) || EG-CNNArea under Receiver Operating Characteristic (AUROC) || MTNLArea under Receiver Operating Characteristic (AUROC) || R2HP
Category (Amazon) || Clothing, Shoes & Jewelry0.5480.5360.5620.601*0.5790.550.5830.5890.623 (+0.022)
Category (Amazon) || Electronics0.5830.5490.5880.616*0.5770.580.6110.6130.661 (+0.045)
Category (Amazon) || Grocery & Gourmet Food0.5360.5260.5530.6020.5320.5460.6110.626*0.657 (+0.031)
Category (Amazon) || Health & Personal Care0.5580.5230.5590.610.5910.5620.6130.620*0.683 (+0.063)
Category (Amazon) || Home & Kitchen0.5680.5370.5650.5970.5690.5730.6030.610*0.646 (+0.036)
Category (Amazon) || Movies & TV0.6030.5580.6210.6340.6030.6070.6480.652*0.713 (+0.061)
Category (Amazon) || Pet Supplies0.560.5420.5850.6030.5480.5580.580.629*0.692 (+0.063)
Category (Amazon) || Tools & Home Improvement0.5840.5580.580.5920.5750.5860.6170.624*0.672 (+0.048)
Category (Amazon) || MACRO AVERAGE0.5680.5410.5770.6070.5720.570.6080.620*0.668 (+0.048)
Category (Amazon) || MICRO AVERAGE (Primary)0.5710.5430.580.6090.5730.5740.6140.625*0.675 (+0.049)
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results of performance (AUROC) of up-to-date methods on identifying helpful reviews evaluated by the test sets of Amazon. (italic f onts$\\ast$: the best performance among the baseline approaches; bold fonts: the state-of-the-art performance of all the approaches)}\n\\begin{tabular}{lrrrlrrrll}\n\\toprule\n & Area under Receiver Operating Characteristic (AUROC) $\\parallel$ STR & Area under Receiver Operating Characteristic (AUROC) $\\parallel$ LEX & Area under Receiver Operating Characteristic (AUROC) $\\parallel$ GALC & Area under Receiver Operating Characteristic (AUROC) $\\parallel$ INQUIRER & Area under Receiver Operating Characteristic (AUROC) $\\parallel$ FUSION (SVM) & Area under Receiver Operating Characteristic (AUROC) $\\parallel$ FUSION (R.F.) & Area under Receiver Operating Characteristic (AUROC) $\\parallel$ EG-CNN & Area under Receiver Operating Characteristic (AUROC) $\\parallel$ MTNL & Area under Receiver Operating Characteristic (AUROC) $\\parallel$ R2HP \\\\\n\\midrule\nCategory (Amazon) $\\parallel$ Clothing, Shoes \\& Jewelry & 0.548 & 0.536 & 0.562 & 0.601* & 0.579 & 0.55 & 0.583 & 0.589 & 0.623 (+0.022) \\\\\nCategory (Amazon) $\\parallel$ Electronics & 0.583 & 0.549 & 0.588 & 0.616* & 0.577 & 0.58 & 0.611 & 0.613 & 0.661 (+0.045) \\\\\nCategory (Amazon) $\\parallel$ Grocery \\& Gourmet Food & 0.536 & 0.526 & 0.553 & 0.602 & 0.532 & 0.546 & 0.611 & 0.626* & 0.657 (+0.031) \\\\\nCategory (Amazon) $\\parallel$ Health \\& Personal Care & 0.558 & 0.523 & 0.559 & 0.61 & 0.591 & 0.562 & 0.613 & 0.620* & 0.683 (+0.063) \\\\\nCategory (Amazon) $\\parallel$ Home \\& Kitchen & 0.568 & 0.537 & 0.565 & 0.597 & 0.569 & 0.573 & 0.603 & 0.610* & 0.646 (+0.036) \\\\\nCategory (Amazon) $\\parallel$ Movies \\& TV & 0.603 & 0.558 & 0.621 & 0.634 & 0.603 & 0.607 & 0.648 & 0.652* & 0.713 (+0.061) \\\\\nCategory (Amazon) $\\parallel$ Pet Supplies & 0.56 & 0.542 & 0.585 & 0.603 & 0.548 & 0.558 & 0.58 & 0.629* & 0.692 (+0.063) \\\\\nCategory (Amazon) $\\parallel$ Tools \\& Home Improvement & 0.584 & 0.558 & 0.58 & 0.592 & 0.575 & 0.586 & 0.617 & 0.624* & 0.672 (+0.048) \\\\\nCategory (Amazon) $\\parallel$ MACRO AVERAGE & 0.568 & 0.541 & 0.577 & 0.607 & 0.572 & 0.57 & 0.608 & 0.620* & 0.668 (+0.048) \\\\\nCategory (Amazon) $\\parallel$ MICRO AVERAGE (Primary) & 0.571 & 0.543 & 0.58 & 0.609 & 0.573 & 0.574 & 0.614 & 0.625* & 0.675 (+0.049) \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Results of performance (AUROC) of up-to-date methods on identifying helpful reviews evaluated by the test sets of Amazon. (italic f onts\u2217: the best performance among the baseline approaches; bold fonts: the state-of-the-art performance of all the approaches)\n
\n \n Area under Receiver Operating Characteristic (AUROC) || STR\n \n Area under Receiver Operating Characteristic (AUROC) || LEX\n \n Area under Receiver Operating Characteristic (AUROC) || GALC\n \n Area under Receiver Operating Characteristic (AUROC) || INQUIRER\n \n Area under Receiver Operating Characteristic (AUROC) || FUSION (SVM)\n \n Area under Receiver Operating Characteristic (AUROC) || FUSION (R.F.)\n \n Area under Receiver Operating Characteristic (AUROC) || EG-CNN\n \n Area under Receiver Operating Characteristic (AUROC) || MTNL\n \n Area under Receiver Operating Characteristic (AUROC) || R2HP\n
\n Category (Amazon) || Clothing, Shoes & Jewelry\n \n 0.548\n \n 0.536\n \n 0.562\n \n 0.601*\n \n 0.579\n \n 0.55\n \n 0.583\n \n 0.589\n \n 0.623 (+0.022)\n
\n Category (Amazon) || Electronics\n \n 0.583\n \n 0.549\n \n 0.588\n \n 0.616*\n \n 0.577\n \n 0.58\n \n 0.611\n \n 0.613\n \n 0.661 (+0.045)\n
\n Category (Amazon) || Grocery & Gourmet Food\n \n 0.536\n \n 0.526\n \n 0.553\n \n 0.602\n \n 0.532\n \n 0.546\n \n 0.611\n \n 0.626*\n \n 0.657 (+0.031)\n
\n Category (Amazon) || Health & Personal Care\n \n 0.558\n \n 0.523\n \n 0.559\n \n 0.61\n \n 0.591\n \n 0.562\n \n 0.613\n \n 0.620*\n \n 0.683 (+0.063)\n
\n Category (Amazon) || Home & Kitchen\n \n 0.568\n \n 0.537\n \n 0.565\n \n 0.597\n \n 0.569\n \n 0.573\n \n 0.603\n \n 0.610*\n \n 0.646 (+0.036)\n
\n Category (Amazon) || Movies & TV\n \n 0.603\n \n 0.558\n \n 0.621\n \n 0.634\n \n 0.603\n \n 0.607\n \n 0.648\n \n 0.652*\n \n 0.713 (+0.061)\n
\n Category (Amazon) || Pet Supplies\n \n 0.56\n \n 0.542\n \n 0.585\n \n 0.603\n \n 0.548\n \n 0.558\n \n 0.58\n \n 0.629*\n \n 0.692 (+0.063)\n
\n Category (Amazon) || Tools & Home Improvement\n \n 0.584\n \n 0.558\n \n 0.58\n \n 0.592\n \n 0.575\n \n 0.586\n \n 0.617\n \n 0.624*\n \n 0.672 (+0.048)\n
\n Category (Amazon) || MACRO AVERAGE\n \n 0.568\n \n 0.541\n \n 0.577\n \n 0.607\n \n 0.572\n \n 0.57\n \n 0.608\n \n 0.620*\n \n 0.668 (+0.048)\n
\n Category (Amazon) || MICRO AVERAGE (Primary)\n \n 0.571\n \n 0.543\n \n 0.58\n \n 0.609\n \n 0.573\n \n 0.574\n \n 0.614\n \n 0.625*\n \n 0.675 (+0.049)\n
\n" }, { "instance_id": "9324086e-657f-405c-8b56-c272437aadb1", "table_id_paper": "D19-1184table_3", "caption": "Results for next utterance retrieval on the Ubuntu dialog corpus. This table shows previous work, and experimental results with two underlying architectures: a dual encoder model and Deep Attention Matching networks. The results shown in the DAM experiments section are performed with the open-sourced implementation of Zhou et al. (2018), which obtains slightly worse performance than they report. All bold-face results are statistically significant to p < 0.01.", "row_header_level": 3, "row_headers": [ [ "Model Name", "Previous Research", "Dual Encoder (Lowe et al., 2015)" ], [ "Model Name", "Previous Research", "MV-LSTM (Pang et al., 2016)" ], [ "Model Name", "Previous Research", "Match-LSTM (Wang and Jiang, 2016)" ], [ "Model Name", "Previous Research", "Multiview (Zhou et al., 2016)" ], [ "Model Name", "Previous Research", "DL2R (Yan et al., 2016)" ], [ "Model Name", "Previous Research", "SMN (Wu et al., 2016)" ], [ "Model Name", "Previous Research", "DAM (Zhou et al., 2018)" ], [ "Model Name", "Dual Encoder Experiments", "Dual Encoder (Lowe et al., 2015)" ], [ "Model Name", "Dual Encoder Experiments", "Ensemble (5)" ], [ "Model Name", "Dual Encoder Experiments", "Multi-Granularity (5)" ], [ "Model Name", "Deep Attention Matching Experiments", "DAM (Zhou et al., 2018) (re-trained)" ], [ "Model Name", "Deep Attention Matching Experiments", "Ensemble (5)" ], [ "Model Name", "Deep Attention Matching Experiments", "Multi-Granularity (5)" ] ], "column_header_level": 1, "column_headers": [ [ "MRR" ], [ "R10@1" ], [ "R2@1" ] ], "contents": [ [ "-", "63.8", "90.1" ], [ "-", "65.3", "90.6" ], [ "", "65.3", "90.4" ], [ "-", "66.2", "90.8" ], [ "-", "62.6", "89.9" ], [ "-", "72.6", "92.6" ], [ "-", "76.7", "93.8" ], [ "76.84", "63.6", "90.9" ], [ "78.91", "66.9", "91.7" ], [ "80.10", "68.7", "91.9" ], [ "83.74", "74.54", "93.08" ], [ "84.03", "74.95", "93.27" ], [ "84.26", "75.3", "93.45" ] ], "metrics_loc": "column", "metrics_type": "['MRR', 'R10@1', 'R2@1']", "target_entity": "['Multi-Granularity (5)']", "table_html": "
MRR R10@1 R2@1
Model Name || Previous Research || Dual Encoder (Lowe et al., 2015) - 63.8 90.1
Model Name || Previous Research || MV-LSTM (Pang et al., 2016) - 65.3 90.6
Model Name || Previous Research || Match-LSTM (Wang and Jiang, 2016) 65.3 90.4
Model Name || Previous Research || Multiview (Zhou et al., 2016) - 66.2 90.8
Model Name || Previous Research || DL2R (Yan et al., 2016) - 62.6 89.9
Model Name || Previous Research || SMN (Wu et al., 2016) - 72.6 92.6
Model Name || Previous Research || DAM (Zhou et al., 2018) - 76.7 93.8
Model Name || Dual Encoder Experiments || Dual Encoder (Lowe et al., 2015) 76.84 63.6 90.9
Model Name || Dual Encoder Experiments || Ensemble (5) 78.91 66.9 91.7
Model Name || Dual Encoder Experiments || Multi-Granularity (5) 80.10 68.7 91.9
Model Name || Deep Attention Matching Experiments || DAM (Zhou et al., 2018) (re-trained) 83.74 74.54 93.08
Model Name || Deep Attention Matching Experiments || Ensemble (5) 84.03 74.95 93.27
Model Name || Deep Attention Matching Experiments || Multi-Granularity (5) 84.26 75.3 93.45
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D19-1184", "page_no": 7, "dir": "emnlp2019", "description": "MGT is applied on top of the dual encoder baseline (Lowe et al., 2015) and Deep Attention Matching networks (Zhou et al., 2018). The results shown in Table 3 show the performance of MGT using two different underlying architectures, as well as previous work. Across both base architectures, MGT outperforms ensembling. The primary difference between these two methods is that MGT explicitly ensures that several granularities of representation are learned. As such, these results reaffirm the hypothesis that learning multiple granularities of representation leads to more diverse models, and more general representations of dialog.", "class_sentence": "[2, 1, 1, 1, 2]", "sentences": "['MGT is applied on top of the dual encoder baseline (Lowe et al., 2015) and Deep Attention Matching networks (Zhou et al., 2018).', 'The results shown in Table 3 show the performance of MGT using two different underlying architectures, as well as previous work.', 'Across both base architectures, MGT outperforms ensembling.', 'The primary difference between these two methods is that MGT explicitly ensures that several granularities of representation are learned.', 'As such, these results reaffirm the hypothesis that learning multiple granularities of representation leads to more diverse models, and more general representations of dialog.']", "header_mention": "[['Multi-Granularity (5)', 'Dual Encoder Experiments', 'Deep Attention Matching Experiments'], ['Multi-Granularity (5)'], ['Multi-Granularity (5)', 'Ensemble (5)'], ['Multi-Granularity (5)', 'Ensemble (5)'], None]", "valid": 1, "image_id": "D19-1184-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results for next utterance retrieval on the Ubuntu dialog corpus. This table shows previous work, and experimental results with two underlying architectures: a dual encoder model and Deep Attention Matching networks. The results shown in the DAM experiments section are performed with the open-sourced implementation of Zhou et al. (2018), which obtains slightly worse performance than they report. All bold-face results are statistically significant to p < 0.01.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n MRRR10@1R2@1
Model Name || Previous Research || Dual Encoder (Lowe et al., 2015)-63.890.1
Model Name || Previous Research || MV-LSTM (Pang et al., 2016)-65.390.6
Model Name || Previous Research || Match-LSTM (Wang and Jiang, 2016)\n 65.390.4
Model Name || Previous Research || Multiview (Zhou et al., 2016)-66.290.8
Model Name || Previous Research || DL2R (Yan et al., 2016)-62.689.9
Model Name || Previous Research || SMN (Wu et al., 2016)-72.692.6
Model Name || Previous Research || DAM (Zhou et al., 2018)-76.793.8
Model Name || Dual Encoder Experiments || Dual Encoder (Lowe et al., 2015)76.8463.690.9
Model Name || Dual Encoder Experiments || Ensemble (5)78.9166.991.7
Model Name || Dual Encoder Experiments || Multi-Granularity (5)80.1068.791.9
Model Name || Deep Attention Matching Experiments || DAM (Zhou et al., 2018) (re-trained)83.7474.5493.08
Model Name || Deep Attention Matching Experiments || Ensemble (5)84.0374.9593.27
Model Name || Deep Attention Matching Experiments || Multi-Granularity (5)84.2675.393.45
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results for next utterance retrieval on the Ubuntu dialog corpus. This table shows previous work, and experimental results with two underlying architectures: a dual encoder model and Deep Attention Matching networks. The results shown in the DAM experiments section are performed with the open-sourced implementation of Zhou et al. (2018), which obtains slightly worse performance than they report. All bold-face results are statistically significant to p $<$ 0.01.}\n\\begin{tabular}{llrr}\n\\toprule\n & MRR & R10@1 & R2@1 \\\\\n\\midrule\nModel Name $\\parallel$ Previous Research $\\parallel$ Dual Encoder (Lowe et al., 2015) & - & 63.8 & 90.1 \\\\\nModel Name $\\parallel$ Previous Research $\\parallel$ MV-LSTM (Pang et al., 2016) & - & 65.3 & 90.6 \\\\\nModel Name $\\parallel$ Previous Research $\\parallel$ Match-LSTM (Wang and Jiang, 2016) & & 65.3 & 90.4 \\\\\nModel Name $\\parallel$ Previous Research $\\parallel$ Multiview (Zhou et al., 2016) & - & 66.2 & 90.8 \\\\\nModel Name $\\parallel$ Previous Research $\\parallel$ DL2R (Yan et al., 2016) & - & 62.6 & 89.9 \\\\\nModel Name $\\parallel$ Previous Research $\\parallel$ SMN (Wu et al., 2016) & - & 72.6 & 92.6 \\\\\nModel Name $\\parallel$ Previous Research $\\parallel$ DAM (Zhou et al., 2018) & - & 76.7 & 93.8 \\\\\nModel Name $\\parallel$ Dual Encoder Experiments $\\parallel$ Dual Encoder (Lowe et al., 2015) & 76.84 & 63.6 & 90.9 \\\\\nModel Name $\\parallel$ Dual Encoder Experiments $\\parallel$ Ensemble (5) & 78.91 & 66.9 & 91.7 \\\\\nModel Name $\\parallel$ Dual Encoder Experiments $\\parallel$ Multi-Granularity (5) & 80.10 & 68.7 & 91.9 \\\\\nModel Name $\\parallel$ Deep Attention Matching Experiments $\\parallel$ DAM (Zhou et al., 2018) (re-trained) & 83.74 & 74.54 & 93.08 \\\\\nModel Name $\\parallel$ Deep Attention Matching Experiments $\\parallel$ Ensemble (5) & 84.03 & 74.95 & 93.27 \\\\\nModel Name $\\parallel$ Deep Attention Matching Experiments $\\parallel$ Multi-Granularity (5) & 84.26 & 75.3 & 93.45 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Results for next utterance retrieval on the Ubuntu dialog corpus. This table shows previous work, and experimental results with two underlying architectures: a dual encoder model and Deep Attention Matching networks. The results shown in the DAM experiments section are performed with the open-sourced implementation of Zhou et al. (2018), which obtains slightly worse performance than they report. All bold-face results are statistically significant to p < 0.01.\n
\n \n MRR\n \n R10@1\n \n R2@1\n
\n Model Name || Previous Research || Dual Encoder (Lowe et al., 2015)\n \n -\n \n 63.8\n \n 90.1\n
\n Model Name || Previous Research || MV-LSTM (Pang et al., 2016)\n \n -\n \n 65.3\n \n 90.6\n
\n Model Name || Previous Research || Match-LSTM (Wang and Jiang, 2016)\n \n \n 65.3\n \n 90.4\n
\n Model Name || Previous Research || Multiview (Zhou et al., 2016)\n \n -\n \n 66.2\n \n 90.8\n
\n Model Name || Previous Research || DL2R (Yan et al., 2016)\n \n -\n \n 62.6\n \n 89.9\n
\n Model Name || Previous Research || SMN (Wu et al., 2016)\n \n -\n \n 72.6\n \n 92.6\n
\n Model Name || Previous Research || DAM (Zhou et al., 2018)\n \n -\n \n 76.7\n \n 93.8\n
\n Model Name || Dual Encoder Experiments || Dual Encoder (Lowe et al., 2015)\n \n 76.84\n \n 63.6\n \n 90.9\n
\n Model Name || Dual Encoder Experiments || Ensemble (5)\n \n 78.91\n \n 66.9\n \n 91.7\n
\n Model Name || Dual Encoder Experiments || Multi-Granularity (5)\n \n 80.10\n \n 68.7\n \n 91.9\n
\n Model Name || Deep Attention Matching Experiments || DAM (Zhou et al., 2018) (re-trained)\n \n 83.74\n \n 74.54\n \n 93.08\n
\n Model Name || Deep Attention Matching Experiments || Ensemble (5)\n \n 84.03\n \n 74.95\n \n 93.27\n
\n Model Name || Deep Attention Matching Experiments || Multi-Granularity (5)\n \n 84.26\n \n 75.3\n \n 93.45\n
\n" }, { "instance_id": "e00f2d1f-ba0b-4226-855e-367d9f76bc74", "table_id_paper": "D19-1315table_6", "caption": "Comparison of the decoder information sharing methods and encoder sharing methods for the job advertisement dataset. The metrics are the same as in Table 3. The proposed method (adopting HCL) achieved the best scores (bold) compared to the other sharing methods.", "row_header_level": 1, "row_headers": [ [ "Baseline (Pointer-Generator Network)" ], [ "Proposed (MTL + SD + HCL)" ], [ "Comparison of Decoder Information Sharing Method MTL + SD" ], [ "Comparison of Decoder Information Sharing Method MTL + SD + Cascade Model" ], [ "Comparison of Decoder Information Sharing Method MTL + SD + Cascade Model (Gold)" ], [ "Comparison of Decoder Information Sharing Method MTL + SD + Soft-Parameter Sharing" ], [ "Comparison of Decoder Information Sharing Method MTL + SD + Non-Hierarchical Consistency Loss" ], [ "Comparison of Decoder Information Sharing Method MTL + SD + HCL with Normalized Attention Weights" ], [ "Comparison of Encoder Information Sharing Method HCL (SD and MTL are not applied)" ], [ "Comparison of Encoder Information Sharing Method SD + HCL (MTL is not applied)" ] ], "column_header_level": 2, "column_headers": [ [ "Headline Generation", "R-1" ], [ " Headline Generation", "R-2" ], [ "Headline Generation", "R-L" ], [ "Key Phrase Generation", "R-1" ], [ "Key Phrase Generation", "R-2" ], [ "Key Phrase Generation", "R-L" ], [ " Classification", "Accuracy" ] ], "contents": [ [ "25.1", "5.3", "21.1", "30.9", "10.6", "28.7", "62.8" ], [ "26.9", "6.1", "22.4", "32.8", "11.2", "30.5", "64.4" ], [ "26.3", "6", "21.8", "32.3", "10.4", "29.9", "63.9" ], [ "26.3", "5.6", "21.6", "31.8", "10.6", "29.5", "64.4" ], [ "26.5", "5.8", "21.9", "32.8", "10.4", "30.3", "64.5" ], [ "25.8", "5.9", "21.4", "32.1", "10", "29.6", "64" ], [ "25.9", "6", "21.4", "32.6", "10.9", "30.2", "64" ], [ "26.2", "6", "21.7", "31.9", "10.5", "29.5", "63.9" ], [ "25.8", "5.6", "21.2", "31", "10.1", "28.7", "63.1" ], [ "25.6", "5.6", "21.5", "31.2", "10.2", "28.9", "62.6" ] ], "metrics_loc": "column", "metrics_type": "['R-1', 'R-2', 'R-L', 'R-1', 'R-2', 'R-L', 'Accuracy']", "target_entity": "['Proposed (MTL + SD + HCL)']", "table_html": "
Headline Generation || R-1 Headline Generation || R-2 Headline Generation || R-L Key Phrase Generation || R-1 Key Phrase Generation || R-2 Key Phrase Generation || R-L Classification || Accuracy
Baseline (Pointer-Generator Network) 25.1 5.3 21.1 30.9 10.6 28.7 62.8
Proposed (MTL + SD + HCL) 26.9 6.1 22.4 32.8 11.2 30.5 64.4
Comparison of Decoder Information Sharing Method MTL + SD 26.3 6 21.8 32.3 10.4 29.9 63.9
Comparison of Decoder Information Sharing Method MTL + SD + Cascade Model 26.3 5.6 21.6 31.8 10.6 29.5 64.4
Comparison of Decoder Information Sharing Method MTL + SD + Cascade Model (Gold) 26.5 5.8 21.9 32.8 10.4 30.3 64.5
Comparison of Decoder Information Sharing Method MTL + SD + Soft-Parameter Sharing 25.8 5.9 21.4 32.1 10 29.6 64
Comparison of Decoder Information Sharing Method MTL + SD + Non-Hierarchical Consistency Loss 25.9 6 21.4 32.6 10.9 30.2 64
Comparison of Decoder Information Sharing Method MTL + SD + HCL with Normalized Attention Weights 26.2 6 21.7 31.9 10.5 29.5 63.9
Comparison of Encoder Information Sharing Method HCL (SD and MTL are not applied) 25.8 5.6 21.2 31 10.1 28.7 63.1
Comparison of Encoder Information Sharing Method SD + HCL (MTL is not applied) 25.6 5.6 21.5 31.2 10.2 28.9 62.6
", "table_name": "Table 6", "table_id": "table_6", "paper_id": "D19-1315", "page_no": 8, "dir": "emnlp2019", "description": "It can be observed from Table 6 that the cascade model achieves a lower score than MTL + SD. However, the cascade model (gold) improves the ROUGE-L score. This result indicates that the classification error propagates to the generators when the cascade model is applied. Our proposed HCL has an advantage in that it does not suffer from an error of the classifier, and thus this method achieves the best score. Furthermore, we compare our proposed HCL model with other settings. A soft-parameter sharing method (Guo et al., 2018) penalizes the difference between parameters in pairs of decoders. The non-hierarchical consistency loss is almost the same as the hierarchical consistency loss (Eqn. 10). We replace a ramp function in Eqn. 10 with an absolute value function. Neither the soft-parameter sharing method nor the non-hierarchical consistency loss has the ability to consider the hierarchy among tasks. It can be observed from Table 6 that both methods achieved a lower score than the hierarchical consistency loss. This is because our hierarchical consistency loss enables the model to penalize a multi-task model with hierarchy among the tasks. Although the HCL with normalized-attention weights adopts the hierarchical consistency loss model indicated in Eqn. 10, however, we substitute the normalized attention weights ?ds ij for non-normalized attention weights eds ij . HCL with normalized attention weights is not as effective as HCL with non-normalized attention weights. Our HCL is based on the assumption that if one specific input word is important for both the shorter and longer text generation tasks, the attention weights of the words for the shorter text generation task would be smaller than the attention weights for the longer text generation task. However, the distributions of the normalized attention weights converge to a few words for the shorter text generation task; thus the normalized attention weights does not satisfy the assumption. As a result, our proposed HCL with non-normalized attention weights can accurately compute this inconsistency, contrary to the HCL with normalized attention weights. Comparison of encoder information sharing methods. To determine the dependence of HCL on MTL, we conduct two experiments. HCL applies the hierarchical consistency loss without MTL and SD, while SD + HCL applies the scheduling sampling and hierarchical consistency loss; however, multi-task learning is not applied. Table 6 presents the results of two methods for the job advertisement dataset.", "class_sentence": "[1, 1, 2, 1, 2, 2, 1, 1, 1, 1, 2, 2, 1, 2, 2, 1, 2, 2, 2, 1]", "sentences": "['It can be observed from Table 6 that the cascade model achieves a lower score than MTL + SD.', 'However, the cascade model (gold) improves the ROUGE-L score.', 'This result indicates that the classification error propagates to the generators when the cascade model is applied.', 'Our proposed HCL has an advantage in that it does not suffer from an error of the classifier, and thus this method achieves the best score.', 'Furthermore, we compare our proposed HCL model with other settings.', 'A soft-parameter sharing method (Guo et al., 2018) penalizes the difference between parameters in pairs of decoders.', 'The non-hierarchical consistency loss is almost the same as the hierarchical consistency loss (Eqn. 10).', 'We replace a ramp function in Eqn. 10 with an absolute value function.', 'Neither the soft-parameter sharing method nor the non-hierarchical consistency loss has the ability to consider the hierarchy among tasks.', 'It can be observed from Table 6 that both methods achieved a lower score than the hierarchical consistency loss.', 'This is because our hierarchical consistency loss enables the model to penalize a multi-task model with hierarchy among the tasks.', 'Although the HCL with normalized-attention weights adopts the hierarchical consistency loss model indicated in Eqn. 10, however, we substitute the normalized attention weights ?ds ij for non-normalized attention weights eds ij .', 'HCL with normalized attention weights is not as effective as HCL with non-normalized attention weights.', 'Our HCL is based on the assumption that if one specific input word is important for both the shorter and longer text generation tasks, the attention weights of the words for the shorter text generation task would be smaller than the attention weights for the longer text generation task.', 'However, the distributions of the normalized attention weights converge to a few words for the shorter text generation task; thus the normalized attention weights does not satisfy the assumption.', 'As a result, our proposed HCL with non-normalized attention weights can accurately compute this inconsistency, contrary to the HCL with normalized attention weights.', 'Comparison of encoder information sharing methods.', 'To determine the dependence of HCL on MTL, we conduct two experiments.', 'HCL applies the hierarchical consistency loss without MTL and SD, while SD + HCL applies the scheduling sampling and hierarchical consistency loss; however, multi-task learning is not applied.', 'Table 6 presents the results of two methods for the job advertisement dataset.']", "header_mention": "[['Comparison of Decoder Information Sharing Method MTL + SD + Cascade Model', 'Proposed (MTL + SD + HCL)'], ['Comparison of Decoder Information Sharing Method MTL + SD + Cascade Model (Gold)'], None, ['Proposed (MTL + SD + HCL)'], None, None, None, None, None, ['Comparison of Decoder Information Sharing Method MTL + SD + Soft-Parameter Sharing', 'Comparison of Decoder Information Sharing Method MTL + SD + Non-Hierarchical Consistency Loss'], None, None, ['Comparison of Decoder Information Sharing Method MTL + SD + HCL with Normalized Attention Weights', 'Proposed (MTL + SD + HCL)'], None, None, ['Proposed (MTL + SD + HCL)', 'Comparison of Decoder Information Sharing Method MTL + SD + HCL with Normalized Attention Weights'], None, None, ['Proposed (MTL + SD + HCL)'], None]", "valid": 1, "image_id": "D19-1315-Table6-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison of the decoder information sharing methods and encoder sharing methods for the job advertisement dataset. The metrics are the same as in Table 3. The proposed method (adopting HCL) achieved the best scores (bold) compared to the other sharing methods.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Headline Generation || R-1Headline Generation || R-2Headline Generation || R-LKey Phrase Generation || R-1Key Phrase Generation || R-2Key Phrase Generation || R-LClassification || Accuracy
Baseline (Pointer-Generator Network)25.15.321.130.910.628.762.8
Proposed (MTL + SD + HCL)26.96.122.432.811.230.564.4
Comparison of Decoder Information Sharing Method MTL + SD26.3621.832.310.429.963.9
Comparison of Decoder Information Sharing Method MTL + SD + Cascade Model26.35.621.631.810.629.564.4
Comparison of Decoder Information Sharing Method MTL + SD + Cascade Model (Gold)26.55.821.932.810.430.364.5
Comparison of Decoder Information Sharing Method MTL + SD + Soft-Parameter Sharing25.85.921.432.11029.664
Comparison of Decoder Information Sharing Method MTL + SD + Non-Hierarchical Consistency Loss25.9621.432.610.930.264
Comparison of Decoder Information Sharing Method MTL + SD + HCL with Normalized Attention Weights26.2621.731.910.529.563.9
Comparison of Encoder Information Sharing Method HCL (SD and MTL are not applied)25.85.621.23110.128.763.1
Comparison of Encoder Information Sharing Method SD + HCL (MTL is not applied)25.65.621.531.210.228.962.6
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{5}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison of the decoder information sharing methods and encoder sharing methods for the job advertisement dataset. The metrics are the same as in Table 3. The proposed method (adopting HCL) achieved the best scores (bold) compared to the other sharing methods.}\n\\begin{tabular}{lrrrrrrr}\n\\toprule\n & Headline Generation $\\parallel$ R-1 & Headline Generation $\\parallel$ R-2 & Headline Generation $\\parallel$ R-L & Key Phrase Generation $\\parallel$ R-1 & Key Phrase Generation $\\parallel$ R-2 & Key Phrase Generation $\\parallel$ R-L & Classification $\\parallel$ Accuracy \\\\\n\\midrule\nBaseline (Pointer-Generator Network) & 25.1 & 5.3 & 21.1 & 30.9 & 10.6 & 28.7 & 62.8 \\\\\nProposed (MTL + SD + HCL) & 26.9 & 6.1 & 22.4 & 32.8 & 11.2 & 30.5 & 64.4 \\\\\nComparison of Decoder Information Sharing Method MTL + SD & 26.3 & 6 & 21.8 & 32.3 & 10.4 & 29.9 & 63.9 \\\\\nComparison of Decoder Information Sharing Method MTL + SD + Cascade Model & 26.3 & 5.6 & 21.6 & 31.8 & 10.6 & 29.5 & 64.4 \\\\\nComparison of Decoder Information Sharing Method MTL + SD + Cascade Model (Gold) & 26.5 & 5.8 & 21.9 & 32.8 & 10.4 & 30.3 & 64.5 \\\\\nComparison of Decoder Information Sharing Method MTL + SD + Soft-Parameter Sharing & 25.8 & 5.9 & 21.4 & 32.1 & 10 & 29.6 & 64 \\\\\nComparison of Decoder Information Sharing Method MTL + SD + Non-Hierarchical Consistency Loss & 25.9 & 6 & 21.4 & 32.6 & 10.9 & 30.2 & 64 \\\\\nComparison of Decoder Information Sharing Method MTL + SD + HCL with Normalized Attention Weights & 26.2 & 6 & 21.7 & 31.9 & 10.5 & 29.5 & 63.9 \\\\\nComparison of Encoder Information Sharing Method HCL (SD and MTL are not applied) & 25.8 & 5.6 & 21.2 & 31 & 10.1 & 28.7 & 63.1 \\\\\nComparison of Encoder Information Sharing Method SD + HCL (MTL is not applied) & 25.6 & 5.6 & 21.5 & 31.2 & 10.2 & 28.9 & 62.6 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 6: Comparison of the decoder information sharing methods and encoder sharing methods for the job advertisement dataset. The metrics are the same as in Table 3. The proposed method (adopting HCL) achieved the best scores (bold) compared to the other sharing methods.\n
\n \n Headline Generation || R-1\n \n Headline Generation || R-2\n \n Headline Generation || R-L\n \n Key Phrase Generation || R-1\n \n Key Phrase Generation || R-2\n \n Key Phrase Generation || R-L\n \n Classification || Accuracy\n
\n Baseline (Pointer-Generator Network)\n \n 25.1\n \n 5.3\n \n 21.1\n \n 30.9\n \n 10.6\n \n 28.7\n \n 62.8\n
\n Proposed (MTL + SD + HCL)\n \n 26.9\n \n 6.1\n \n 22.4\n \n 32.8\n \n 11.2\n \n 30.5\n \n 64.4\n
\n Comparison of Decoder Information Sharing Method MTL + SD\n \n 26.3\n \n 6\n \n 21.8\n \n 32.3\n \n 10.4\n \n 29.9\n \n 63.9\n
\n Comparison of Decoder Information Sharing Method MTL + SD + Cascade Model\n \n 26.3\n \n 5.6\n \n 21.6\n \n 31.8\n \n 10.6\n \n 29.5\n \n 64.4\n
\n Comparison of Decoder Information Sharing Method MTL + SD + Cascade Model (Gold)\n \n 26.5\n \n 5.8\n \n 21.9\n \n 32.8\n \n 10.4\n \n 30.3\n \n 64.5\n
\n Comparison of Decoder Information Sharing Method MTL + SD + Soft-Parameter Sharing\n \n 25.8\n \n 5.9\n \n 21.4\n \n 32.1\n \n 10\n \n 29.6\n \n 64\n
\n Comparison of Decoder Information Sharing Method MTL + SD + Non-Hierarchical Consistency Loss\n \n 25.9\n \n 6\n \n 21.4\n \n 32.6\n \n 10.9\n \n 30.2\n \n 64\n
\n Comparison of Decoder Information Sharing Method MTL + SD + HCL with Normalized Attention Weights\n \n 26.2\n \n 6\n \n 21.7\n \n 31.9\n \n 10.5\n \n 29.5\n \n 63.9\n
\n Comparison of Encoder Information Sharing Method HCL (SD and MTL are not applied)\n \n 25.8\n \n 5.6\n \n 21.2\n \n 31\n \n 10.1\n \n 28.7\n \n 63.1\n
\n Comparison of Encoder Information Sharing Method SD + HCL (MTL is not applied)\n \n 25.6\n \n 5.6\n \n 21.5\n \n 31.2\n \n 10.2\n \n 28.9\n \n 62.6\n
\n" }, { "instance_id": "8130915e-ce03-4c27-82ff-a0915665c6df", "table_id_paper": "D19-1381table_4", "caption": "Nested and overlapping event detection F1 (%) score performance on the CG task 2013 development set.", "row_header_level": 2, "row_headers": [ [ "Model", "TEES" ], [ "Model", "SBNN k = 8" ] ], "column_header_level": 1, "column_headers": [ [ "Nested" ], [ "Overlapping" ], [ "Flat" ], [ "Overall F1 (%)" ] ], "contents": [ [ "42.7", "34.49", "56.81", "52.16" ], [ "45.24", "36.92", "60.5", "54.36" ] ], "metrics_loc": "column", "metrics_type": "['F1', 'F1', 'F1', 'F1']", "target_entity": "['SBNN k = 8']", "table_html": "
Nested Overlapping Flat Overall F1 (%)
Model || TEES 42.7 34.49 56.81 52.16
Model || SBNN k = 8 45.24 36.92 60.5 54.36
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D19-1381", "page_no": 5, "dir": "emnlp2019", "description": "Table 4 shows the performance comparison of the models on nested, overlapping and flat event detection. Our model yields higher F1-scores than TEES which can be attributed to its ability to maintain multiple beams and to detect events from all these beams during search.", "class_sentence": "[1, 1]", "sentences": "['Table 4 shows the performance comparison of the models on nested, overlapping and flat event detection.', 'Our model yields higher F1-scores than TEES which can be attributed to its ability to maintain multiple beams and to detect events from all these beams during search.']", "header_mention": "[['Nested', 'Overlapping', 'Flat'], ['SBNN k = 8', 'TEES', 'Overall F1 (%)']]", "valid": 1, "image_id": "D19-1381-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Nested and overlapping event detection F1 (%) score performance on the CG task 2013 development set.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n NestedOverlappingFlatOverall F1 (%)
Model || TEES42.734.4956.8152.16
Model || SBNN k = 845.2436.9260.554.36
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Nested and overlapping event detection F1 (\\%) score performance on the CG task 2013 development set.}\n\\begin{tabular}{lrrrr}\n\\toprule\n & Nested & Overlapping & Flat & Overall F1 (\\%) \\\\\n\\midrule\nModel $\\parallel$ TEES & 42.7 & 34.49 & 56.81 & 52.16 \\\\\nModel $\\parallel$ SBNN k $=$ 8 & 45.24 & 36.92 & 60.5 & 54.36 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Nested and overlapping event detection F1 (%) score performance on the CG task 2013 development set.\n
\n \n Nested\n \n Overlapping\n \n Flat\n \n Overall F1 (%)\n
\n Model || TEES\n \n 42.7\n \n 34.49\n \n 56.81\n \n 52.16\n
\n Model || SBNN k = 8\n \n 45.24\n \n 36.92\n \n 60.5\n \n 54.36\n
\n" }, { "instance_id": "a4384587-be59-49eb-bf19-88c283eda4ab", "table_id_paper": "D19-1399table_4", "caption": "Performance comparison on the OntoNotes 5.0 Chinese Dataset.", "row_header_level": 3, "row_headers": [ [ "Model", "Model", "Pradhan et al. (2013)" ], [ "Model", "Model", "Lattice LSTM (Z&Y, 2018)" ], [ "Model", "Model", "BiLSTM-CRF (L = 0)" ], [ "Model", "Model", "BiLSTM-CRF (L = 1)" ], [ "Model", "Model", "BiLSTM-CRF (L = 2)" ], [ "Model", "Model", "BiLSTM-CRF (L = 3)" ], [ "Model", "Model", "BiLSTM-GCN-CRF" ], [ "Model", "Model", "DGLSTM-CRF (L = 0)" ], [ "Model", "Model", "DGLSTM-CRF (L = 1)" ], [ "Model", "Model", "DGLSTM-CRF (L = 2)" ], [ "Model", "Model", "DGLSTM-CRF (L = 3)" ], [ "Model", "Contextualized Word Representation", "BiLSTM-CRF (L = 0) + ELMo" ], [ "Model", "Contextualized Word Representation", "BiLSTM-CRF (L = 1) + ELMo" ], [ "Model", "Contextualized Word Representation", "BiLSTM-CRF(L = 2) + ELMo" ], [ "Model", "Contextualized Word Representation", "BiLSTM-CRF (L = 3) + ELMo" ], [ "Model", "Contextualized Word Representation", "BiLSTM-GCN-CRF + ELMo" ], [ "Model", "Contextualized Word Representation", "DGLSTM-CRF (L = 0) + ELMo" ], [ "Model", "Contextualized Word Representation", "DGLSTM-CRF (L = 1) + ELMo" ], [ "Model", "Contextualized Word Representation", "DGLSTM-CRF (L = 2) + ELMo" ], [ "Model", "Contextualized Word Representation", "DGLSTM-CRF (L = 3) + ELMo" ] ], "column_header_level": 1, "column_headers": [ [ "Prec." ], [ "Rec." ], [ "F1" ] ], "contents": [ [ "78.2", "66.45", "71.85" ], [ "76.34", "77.01", "76.67" ], [ "76.67", "67.79", "71.95" ], [ "78.45", "74.59", "76.47" ], [ "77.94", "75.33", "76.61" ], [ "76.17", "75.23", "75.7" ], [ "76.35", "75.89", "76.12" ], [ "76.91", "70.65", "73.65" ], [ "77.79", "75.29", "76.52" ], [ "77.4", "77.41", "77.4" ], [ "77.01", "74.9", "75.94" ], [ "75.2", "73.39", "74.28" ], [ "79.2", "79.21", "79.2" ], [ "78.49", "79.44", "78.96" ], [ "78.54", "79.76", "79.14" ], [ "78.71", "79.29", "79" ], [ "76.27", "74.61", "75.43" ], [ "78.91", "80.22", "79.56" ], [ "78.86", "81", "79.92" ], [ "79.3", "79.86", "79.58" ] ], "metrics_loc": "column", "metrics_type": "['Prec.', 'Rec.', 'F1']", "target_entity": "['BiLSTM-CRF (L = 0)', 'BiLSTM-CRF (L = 1)', 'BiLSTM-CRF (L = 2)', 'BiLSTM-CRF (L = 3)', 'BiLSTM-CRF (L = 0) + ELMo', 'BiLSTM-CRF (L = 1) + ELMo', 'BiLSTM-CRF(L = 2) + ELMo', 'BiLSTM-CRF (L = 3) + ELMo', 'DGLSTM-CRF (L = 0)', 'DGLSTM-CRF (L = 1)', 'DGLSTM-CRF (L = 2)', 'DGLSTM-CRF (L = 3)', 'DGLSTM-CRF (L = 0) + ELMo', 'DGLSTM-CRF (L = 1) + ELMo', 'DGLSTM-CRF (L = 2) + ELMo', 'DGLSTM-CRF (L = 3) + ELMo']", "table_html": "
Prec. Rec. F1
Model || Model || Pradhan et al. (2013) 78.2 66.45 71.85
Model || Model || Lattice LSTM (Z&Y, 2018) 76.34 77.01 76.67
Model || Model || BiLSTM-CRF (L = 0) 76.67 67.79 71.95
Model || Model || BiLSTM-CRF (L = 1) 78.45 74.59 76.47
Model || Model || BiLSTM-CRF (L = 2) 77.94 75.33 76.61
Model || Model || BiLSTM-CRF (L = 3) 76.17 75.23 75.7
Model || Model || BiLSTM-GCN-CRF 76.35 75.89 76.12
Model || Model || DGLSTM-CRF (L = 0) 76.91 70.65 73.65
Model || Model || DGLSTM-CRF (L = 1) 77.79 75.29 76.52
Model || Model || DGLSTM-CRF (L = 2) 77.4 77.41 77.4
Model || Model || DGLSTM-CRF (L = 3) 77.01 74.9 75.94
Model || Contextualized Word Representation || BiLSTM-CRF (L = 0) + ELMo 75.2 73.39 74.28
Model || Contextualized Word Representation || BiLSTM-CRF (L = 1) + ELMo 79.2 79.21 79.2
Model || Contextualized Word Representation || BiLSTM-CRF(L = 2) + ELMo 78.49 79.44 78.96
Model || Contextualized Word Representation || BiLSTM-CRF (L = 3) + ELMo 78.54 79.76 79.14
Model || Contextualized Word Representation || BiLSTM-GCN-CRF + ELMo 78.71 79.29 79
Model || Contextualized Word Representation || DGLSTM-CRF (L = 0) + ELMo 76.27 74.61 75.43
Model || Contextualized Word Representation || DGLSTM-CRF (L = 1) + ELMo 78.91 80.22 79.56
Model || Contextualized Word Representation || DGLSTM-CRF (L = 2) + ELMo 78.86 81 79.92
Model || Contextualized Word Representation || DGLSTM-CRF (L = 3) + ELMo 79.3 79.86 79.58
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D19-1399", "page_no": 6, "dir": "emnlp2019", "description": "OntoNotes Chinese Table 4 shows the performance comparison on the Chinese datasets. We compare our models against the state-of-the-art NER model on this dataset, Lattice LSTM (Zhang and Yang, 2018). Our implementation of the strong BiLSTM-CRF baseline achieves comparable performance against the Lattice LSTM. Similar to the English dataset, our model with L = 0 significantly improves the performance compared to the BiLSTM-CRF (L = 0) model. Our DGLSTM-CRF model achieves the best performance with L = 2 and is consistently better (p < 0.02) than the strong BiLSTM-CRF baselines. As we can see from the table, the improvements of the DGLSTM-CRF model mainly come from recall (p < 0.001) compared to the BiLSTM model, especially in the scenario with word embeddings only. Empirically, we also found that those correctly retrieved entities of the DGLSTM-CRF (compared against the baseline) mostly correlate with the following dependency relations: \u201cnn\u201d, \u201cnsubj\u201d, \u201cnummod\u201d. However, DGLSTM-CRF achieves lower precisions against BiLSTM-CRF, which indicates that the DGLSTM-CRF model makes more false-positive predictions. The reason could be the relatively lower ratio of ST(%) as shown in Table 2, which means some of the entities do not form subtrees under the complete dependency trees. In such a scenario, the model may not correctly identify the boundary of the entities, which results in lower precision.", "class_sentence": "[1, 2, 2, 1, 1, 1, 1, 2, 2]", "sentences": "['OntoNotes Chinese Table 4 shows the performance comparison on the Chinese datasets.', 'We compare our models against the state-of-the-art NER model on this dataset, Lattice LSTM (Zhang and Yang, 2018).', 'Our implementation of the strong BiLSTM-CRF baseline achieves comparable performance against the Lattice LSTM.', 'Similar to the English dataset, our model with L = 0 significantly improves the performance compared to the BiLSTM-CRF (L = 0) model.', 'Our DGLSTM-CRF model achieves the best performance with L = 2 and is consistently better (p < 0.02) than the strong BiLSTM-CRF baselines. As we can see from the table, the improvements of the DGLSTM-CRF model mainly come from recall (p < 0.001) compared to the BiLSTM model, especially in the scenario with word embeddings only.', 'Empirically, we also found that those correctly retrieved entities of the DGLSTM-CRF (compared against the baseline) mostly correlate with the following dependency relations: \u201cnn\u201d, \u201cnsubj\u201d, \u201cnummod\u201d.', 'However, DGLSTM-CRF achieves lower precisions against BiLSTM-CRF, which indicates that the DGLSTM-CRF model makes more false-positive predictions.', 'The reason could be the relatively lower ratio of ST(%) as shown in Table 2, which means some of the entities do not form subtrees under the complete dependency trees.', 'In such a scenario, the model may not correctly identify the boundary of the entities, which results in lower precision.']", "header_mention": "[None, ['Lattice LSTM (Z&Y, 2018)'], ['Lattice LSTM (Z&Y, 2018)', 'BiLSTM-CRF (L = 0)', 'BiLSTM-CRF (L = 1)', 'BiLSTM-CRF (L = 2)', 'BiLSTM-CRF (L = 3)', 'BiLSTM-CRF (L = 0) + ELMo', 'BiLSTM-CRF (L = 1) + ELMo', 'BiLSTM-CRF(L = 2) + ELMo', 'BiLSTM-CRF (L = 3) + ELMo'], ['BiLSTM-CRF (L = 0)', 'DGLSTM-CRF (L = 0)'], ['DGLSTM-CRF (L = 2)', 'BiLSTM-CRF (L = 0)', 'BiLSTM-CRF (L = 1)', 'BiLSTM-CRF (L = 2)', 'BiLSTM-CRF (L = 3)', 'BiLSTM-CRF (L = 0) + ELMo', 'BiLSTM-CRF (L = 1) + ELMo', 'BiLSTM-CRF(L = 2) + ELMo', 'BiLSTM-CRF (L = 3) + ELMo'], ['DGLSTM-CRF (L = 0)', 'DGLSTM-CRF (L = 1)', 'DGLSTM-CRF (L = 2)', 'DGLSTM-CRF (L = 3)', 'DGLSTM-CRF (L = 0) + ELMo', 'DGLSTM-CRF (L = 1) + ELMo', 'DGLSTM-CRF (L = 2) + ELMo', 'DGLSTM-CRF (L = 3) + ELMo'], ['BiLSTM-CRF (L = 0)', 'BiLSTM-CRF (L = 1)', 'BiLSTM-CRF (L = 2)', 'BiLSTM-CRF (L = 3)', 'BiLSTM-CRF (L = 0) + ELMo', 'BiLSTM-CRF (L = 1) + ELMo', 'BiLSTM-CRF(L = 2) + ELMo', 'BiLSTM-CRF (L = 3) + ELMo', 'DGLSTM-CRF (L = 0)'], None, None]", "valid": 1, "image_id": "D19-1399-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance comparison on the OntoNotes 5.0 Chinese Dataset.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Prec.Rec.F1
Model || Model || Pradhan et al. (2013)78.266.4571.85
Model || Model || Lattice LSTM (Z&Y, 2018)76.3477.0176.67
Model || Model || BiLSTM-CRF (L = 0)76.6767.7971.95
Model || Model || BiLSTM-CRF (L = 1)78.4574.5976.47
Model || Model || BiLSTM-CRF (L = 2)77.9475.3376.61
Model || Model || BiLSTM-CRF (L = 3)76.1775.2375.7
Model || Model || BiLSTM-GCN-CRF76.3575.8976.12
Model || Model || DGLSTM-CRF (L = 0)76.9170.6573.65
Model || Model || DGLSTM-CRF (L = 1)77.7975.2976.52
Model || Model || DGLSTM-CRF (L = 2)77.477.4177.4
Model || Model || DGLSTM-CRF (L = 3)77.0174.975.94
Model || Contextualized Word Representation || BiLSTM-CRF (L = 0) + ELMo75.273.3974.28
Model || Contextualized Word Representation || BiLSTM-CRF (L = 1) + ELMo79.279.2179.2
Model || Contextualized Word Representation || BiLSTM-CRF(L = 2) + ELMo78.4979.4478.96
Model || Contextualized Word Representation || BiLSTM-CRF (L = 3) + ELMo78.5479.7679.14
Model || Contextualized Word Representation || BiLSTM-GCN-CRF + ELMo78.7179.2979
Model || Contextualized Word Representation || DGLSTM-CRF (L = 0) + ELMo76.2774.6175.43
Model || Contextualized Word Representation || DGLSTM-CRF (L = 1) + ELMo78.9180.2279.56
Model || Contextualized Word Representation || DGLSTM-CRF (L = 2) + ELMo78.868179.92
Model || Contextualized Word Representation || DGLSTM-CRF (L = 3) + ELMo79.379.8679.58
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance comparison on the OntoNotes 5.0 Chinese Dataset.}\n\\begin{tabular}{lrrr}\n\\toprule\n & Prec. & Rec. & F1 \\\\\n\\midrule\nModel $\\parallel$ Model $\\parallel$ Pradhan et al. (2013) & 78.2 & 66.45 & 71.85 \\\\\nModel $\\parallel$ Model $\\parallel$ Lattice LSTM (Z\\&Y, 2018) & 76.34 & 77.01 & 76.67 \\\\\nModel $\\parallel$ Model $\\parallel$ BiLSTM-CRF (L $=$ 0) & 76.67 & 67.79 & 71.95 \\\\\nModel $\\parallel$ Model $\\parallel$ BiLSTM-CRF (L $=$ 1) & 78.45 & 74.59 & 76.47 \\\\\nModel $\\parallel$ Model $\\parallel$ BiLSTM-CRF (L $=$ 2) & 77.94 & 75.33 & 76.61 \\\\\nModel $\\parallel$ Model $\\parallel$ BiLSTM-CRF (L $=$ 3) & 76.17 & 75.23 & 75.7 \\\\\nModel $\\parallel$ Model $\\parallel$ BiLSTM-GCN-CRF & 76.35 & 75.89 & 76.12 \\\\\nModel $\\parallel$ Model $\\parallel$ DGLSTM-CRF (L $=$ 0) & 76.91 & 70.65 & 73.65 \\\\\nModel $\\parallel$ Model $\\parallel$ DGLSTM-CRF (L $=$ 1) & 77.79 & 75.29 & 76.52 \\\\\nModel $\\parallel$ Model $\\parallel$ DGLSTM-CRF (L $=$ 2) & 77.4 & 77.41 & 77.4 \\\\\nModel $\\parallel$ Model $\\parallel$ DGLSTM-CRF (L $=$ 3) & 77.01 & 74.9 & 75.94 \\\\\nModel $\\parallel$ Contextualized Word Representation $\\parallel$ BiLSTM-CRF (L $=$ 0) + ELMo & 75.2 & 73.39 & 74.28 \\\\\nModel $\\parallel$ Contextualized Word Representation $\\parallel$ BiLSTM-CRF (L $=$ 1) + ELMo & 79.2 & 79.21 & 79.2 \\\\\nModel $\\parallel$ Contextualized Word Representation $\\parallel$ BiLSTM-CRF(L $=$ 2) + ELMo & 78.49 & 79.44 & 78.96 \\\\\nModel $\\parallel$ Contextualized Word Representation $\\parallel$ BiLSTM-CRF (L $=$ 3) + ELMo & 78.54 & 79.76 & 79.14 \\\\\nModel $\\parallel$ Contextualized Word Representation $\\parallel$ BiLSTM-GCN-CRF + ELMo & 78.71 & 79.29 & 79 \\\\\nModel $\\parallel$ Contextualized Word Representation $\\parallel$ DGLSTM-CRF (L $=$ 0) + ELMo & 76.27 & 74.61 & 75.43 \\\\\nModel $\\parallel$ Contextualized Word Representation $\\parallel$ DGLSTM-CRF (L $=$ 1) + ELMo & 78.91 & 80.22 & 79.56 \\\\\nModel $\\parallel$ Contextualized Word Representation $\\parallel$ DGLSTM-CRF (L $=$ 2) + ELMo & 78.86 & 81 & 79.92 \\\\\nModel $\\parallel$ Contextualized Word Representation $\\parallel$ DGLSTM-CRF (L $=$ 3) + ELMo & 79.3 & 79.86 & 79.58 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Performance comparison on the OntoNotes 5.0 Chinese Dataset.\n
\n \n Prec.\n \n Rec.\n \n F1\n
\n Model || Model || Pradhan et al. (2013)\n \n 78.2\n \n 66.45\n \n 71.85\n
\n Model || Model || Lattice LSTM (Z&Y, 2018)\n \n 76.34\n \n 77.01\n \n 76.67\n
\n Model || Model || BiLSTM-CRF (L = 0)\n \n 76.67\n \n 67.79\n \n 71.95\n
\n Model || Model || BiLSTM-CRF (L = 1)\n \n 78.45\n \n 74.59\n \n 76.47\n
\n Model || Model || BiLSTM-CRF (L = 2)\n \n 77.94\n \n 75.33\n \n 76.61\n
\n Model || Model || BiLSTM-CRF (L = 3)\n \n 76.17\n \n 75.23\n \n 75.7\n
\n Model || Model || BiLSTM-GCN-CRF\n \n 76.35\n \n 75.89\n \n 76.12\n
\n Model || Model || DGLSTM-CRF (L = 0)\n \n 76.91\n \n 70.65\n \n 73.65\n
\n Model || Model || DGLSTM-CRF (L = 1)\n \n 77.79\n \n 75.29\n \n 76.52\n
\n Model || Model || DGLSTM-CRF (L = 2)\n \n 77.4\n \n 77.41\n \n 77.4\n
\n Model || Model || DGLSTM-CRF (L = 3)\n \n 77.01\n \n 74.9\n \n 75.94\n
\n Model || Contextualized Word Representation || BiLSTM-CRF (L = 0) + ELMo\n \n 75.2\n \n 73.39\n \n 74.28\n
\n Model || Contextualized Word Representation || BiLSTM-CRF (L = 1) + ELMo\n \n 79.2\n \n 79.21\n \n 79.2\n
\n Model || Contextualized Word Representation || BiLSTM-CRF(L = 2) + ELMo\n \n 78.49\n \n 79.44\n \n 78.96\n
\n Model || Contextualized Word Representation || BiLSTM-CRF (L = 3) + ELMo\n \n 78.54\n \n 79.76\n \n 79.14\n
\n Model || Contextualized Word Representation || BiLSTM-GCN-CRF + ELMo\n \n 78.71\n \n 79.29\n \n 79\n
\n Model || Contextualized Word Representation || DGLSTM-CRF (L = 0) + ELMo\n \n 76.27\n \n 74.61\n \n 75.43\n
\n Model || Contextualized Word Representation || DGLSTM-CRF (L = 1) + ELMo\n \n 78.91\n \n 80.22\n \n 79.56\n
\n Model || Contextualized Word Representation || DGLSTM-CRF (L = 2) + ELMo\n \n 78.86\n \n 81\n \n 79.92\n
\n Model || Contextualized Word Representation || DGLSTM-CRF (L = 3) + ELMo\n \n 79.3\n \n 79.86\n \n 79.58\n
\n" }, { "instance_id": "3bac6c72-2f87-4044-810f-195f3c881342", "table_id_paper": "D19-1403table_2", "caption": "Comparison of mean accuracy (%) on ARSC", "row_header_level": 2, "row_headers": [ [ "Model", "Matching Networks (Vinyals et al., 2016)" ], [ "Model", "Prototypical Networks (Snell et al., 2017)" ], [ "Model", "Graph Network (Garcia and Bruna, 2017)" ], [ "Model", "Relation Network (Sung et al., 2018)" ], [ "Model", "SNAIL (Mishra et al., 2018)" ], [ "Model", "ROBUSTTC-FSL (Yu et al., 2018)" ], [ "Model", "Induction Networks (ours)" ] ], "column_header_level": 1, "column_headers": [ [ "Mean Acc" ] ], "contents": [ [ "65.73" ], [ "68.17" ], [ "82.61" ], [ "83.07" ], [ "82.57" ], [ "83.12" ], [ "85.63" ] ], "metrics_loc": "column", "metrics_type": "['Mean Acc']", "target_entity": "['Induction Networks (ours)']", "table_html": "
Mean Acc
Model || Matching Networks (Vinyals et al., 2016) 65.73
Model || Prototypical Networks (Snell et al., 2017) 68.17
Model || Graph Network (Garcia and Bruna, 2017) 82.61
Model || Relation Network (Sung et al., 2018) 83.07
Model || SNAIL (Mishra et al., 2018) 82.57
Model || ROBUSTTC-FSL (Yu et al., 2018) 83.12
Model || Induction Networks (ours) 85.63
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D19-1403", "page_no": 6, "dir": "emnlp2019", "description": "5.3 Experiment Results Overall Performance Experiment results on ARSC are presented in Table 2. The proposed Induction Networks achieves a 85.63% accuracy, outperforming the existing state-of-the-art model, ROBUSTTC-FSL, by a notable 3% improvement. We due the improvement to the fact that ROBUSTTC-FSL builds a general metric method by integrating several metrics at the sample level, which faces the difficulty of getting rid of the noise among different expressions in the same class. In addition to that, the task-clustering-based method used by ROBUSTTC-FSL must be found on the relevance matrix, which is inefficient when applied to real-world scenarios where the tasks change rapidly. Our Induction Networks, however, is trained in the meta-learning framework with more flexible generalization and its induction ability can hence be accumulated through different tasks. Ablation Study To analyze the effect of varying different components of the Induction Module and Relation Module, we further report the ablation experiments on the ARSC dataset as shown in Table 4. We can see that the best performance is achieved when we used 3 iterations, corresponding to the best result reported in Table 2 (more rounds of iterations did not further improve the performance), and the table shows the effectiveness of the routing component. We also changed the Induction Module with sum and self-attention and changed Relation Module with cosine distance. Changes in the performances validate the benefit of both the Relation Module and Induction Module. The Attention+Relation models the induction ability by self-attention mechanism, but the ability is limited by the learnt attention parameters. Conversely, the proposed dynamic routing induction method captures class-level information by automatically adjusting the coupling coefficients according to inputted support sets, which is more suitable for the few-shot learning task.", "class_sentence": "[2, 1, 2, 2, 2, 0, 0, 0, 0, 0, 0]", "sentences": "['5.3 Experiment Results Overall Performance Experiment results on ARSC are presented in Table 2.', 'The proposed Induction Networks achieves a 85.63% accuracy, outperforming the existing state-of-the-art model, ROBUSTTC-FSL, by a notable 3% improvement.', 'We due the improvement to the fact that ROBUSTTC-FSL builds a general metric method by integrating several metrics at the sample level, which faces the difficulty of getting rid of the noise among different expressions in the same class.', 'In addition to that, the task-clustering-based method used by ROBUSTTC-FSL must be found on the relevance matrix, which is inefficient when applied to real-world scenarios where the tasks change rapidly.', 'Our Induction Networks, however, is trained in the meta-learning framework with more flexible generalization and its induction ability can hence be accumulated through different tasks.', 'Ablation Study To analyze the effect of varying different components of the Induction Module and Relation Module, we further report the ablation experiments on the ARSC dataset as shown in Table 4.', 'We can see that the best performance is achieved when we used 3 iterations, corresponding to the best result reported in Table 2 (more rounds of iterations did not further improve the performance), and the table shows the effectiveness of the routing component.', 'We also changed the Induction Module with sum and self-attention and changed Relation Module with cosine distance.', 'Changes in the performances validate the benefit of both the Relation Module and Induction Module.', 'The Attention+Relation models the induction ability by self-attention mechanism, but the ability is limited by the learnt attention parameters.', 'Conversely, the proposed dynamic routing induction method captures class-level information by automatically adjusting the coupling coefficients according to inputted support sets, which is more suitable for the few-shot learning task.']", "header_mention": "[None, ['Induction Networks (ours)', 'Mean Acc', 'ROBUSTTC-FSL (Yu et al., 2018)', 'Relation Network (Sung et al., 2018)', 'SNAIL (Mishra et al., 2018)'], ['ROBUSTTC-FSL (Yu et al., 2018)', 'Relation Network (Sung et al., 2018)', 'SNAIL (Mishra et al., 2018)'], ['ROBUSTTC-FSL (Yu et al., 2018)', 'Relation Network (Sung et al., 2018)', 'SNAIL (Mishra et al., 2018)'], ['Induction Networks (ours)'], None, None, None, None, None, None]", "valid": 1, "image_id": "D19-1403-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison of mean accuracy (%) on ARSC

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Mean Acc
Model || Matching Networks (Vinyals et al., 2016)65.73
Model || Prototypical Networks (Snell et al., 2017)68.17
Model || Graph Network (Garcia and Bruna, 2017)82.61
Model || Relation Network (Sung et al., 2018)83.07
Model || SNAIL (Mishra et al., 2018)82.57
Model || ROBUSTTC-FSL (Yu et al., 2018)83.12
Model || Induction Networks (ours)85.63
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison of mean accuracy (\\%) on ARSC}\n\\begin{tabular}{lr}\n\\toprule\n & Mean Acc \\\\\n\\midrule\nModel $\\parallel$ Matching Networks (Vinyals et al., 2016) & 65.73 \\\\\nModel $\\parallel$ Prototypical Networks (Snell et al., 2017) & 68.17 \\\\\nModel $\\parallel$ Graph Network (Garcia and Bruna, 2017) & 82.61 \\\\\nModel $\\parallel$ Relation Network (Sung et al., 2018) & 83.07 \\\\\nModel $\\parallel$ SNAIL (Mishra et al., 2018) & 82.57 \\\\\nModel $\\parallel$ ROBUSTTC-FSL (Yu et al., 2018) & 83.12 \\\\\nModel $\\parallel$ Induction Networks (ours) & 85.63 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Comparison of mean accuracy (%) on ARSC\n
\n \n Mean Acc\n
\n Model || Matching Networks (Vinyals et al., 2016)\n \n 65.73\n
\n Model || Prototypical Networks (Snell et al., 2017)\n \n 68.17\n
\n Model || Graph Network (Garcia and Bruna, 2017)\n \n 82.61\n
\n Model || Relation Network (Sung et al., 2018)\n \n 83.07\n
\n Model || SNAIL (Mishra et al., 2018)\n \n 82.57\n
\n Model || ROBUSTTC-FSL (Yu et al., 2018)\n \n 83.12\n
\n Model || Induction Networks (ours)\n \n 85.63\n
\n" }, { "instance_id": "b6f945c9-311a-4093-a47e-78c0307b6b66", "table_id_paper": "D19-1450table_2", "caption": "Results of bilingual lexicon induction (accuracy % P@1) for similar and distant language pairs on the dataset BLI-1. Procrustes and MUSE represent the supervised and unsupervised model of Lample et al. (2018), VecMap is the supervised model of Artetxe et al. (2017). Word translations are retrieved by using CSLS. Bold face indicates the best result overall and italics indicate the best result between the two columns without refinement.", "row_header_level": 2, "row_headers": [ [ "Similar language pairs", "de-en" ], [ "Similar language pairs", "en-de" ], [ "Similar language pairs", "es-en" ], [ "Similar language pairs", "en-es" ], [ "Similar language pairs", "fr-en" ], [ "Similar language pairs", "en-fr" ], [ "Similar language pairs", "it-en" ], [ "Similar language pairs", "en-it" ], [ "Metric", "avg." ], [ "Distant language pairs", "fi-en" ], [ "Distant language pairs", "en-fi" ], [ "Distant language pairs", "ru-en" ], [ "Distant language pairs", "en-ru" ], [ "Distant language pairs", "tr-en" ], [ "Distant language pairs", "en-tr" ], [ "Distant language pairs", "zh-en" ], [ "Distant language pairs", "en-zh" ], [ "Metric", "avg." ] ], "column_header_level": 2, "column_headers": [ [ "Supervised", "VecMap" ], [ "Supervised", "Procrustes" ], [ "Without refinement", "MUSE" ], [ "Without refinement", "MUSEMethod" ], [ "With refinement", "Our" ], [ "With refinement", "Method" ] ], "contents": [ [ "72.5", "72", "55.3", "66.3", "72.2", "72.4" ], [ "72.5", "72.1", "59.2", "68.2", "72.5", "73.4" ], [ "83.4", "82.9", "78.1", "78", "83.1", "83.6" ], [ "81.9", "81.5", "75.5", "74.9", "81.1", "80.9" ], [ "81.5", "82.2", "72.2", "75.6", "81.2", "83.3" ], [ "81.3", "81.3", "77.8", "78.3", "82.2", "82.2" ], [ "77.7", "79.1", "61", "63.3", "76.4", "76.2" ], [ "77.2", "77.5", "63.3", "64.3", "77.4", "78.2" ], [ "78.5", "78.6", "67.8", "71.1", "78.3", "78.8" ], [ "55.9", "56.2", "0", "42.8", "28.6", "53.5" ], [ "39.5", "39.3", "0", "35.2", "21", "43" ], [ "60.5", "61.3", "43.8", "54.3", "49.9", "58.9" ], [ "48", "50.9", "28.7", "43.2", "37.6", "51.1" ], [ "55.7", "58", "12.2", "42.1", "25.5", "54.7" ], [ "37.3", "37.1", "26.2", "33.1", "39.4", "41.4" ], [ "45", "41.2", "26.9", "32.1", "30.8", "42" ], [ "45.4", "52.1", "29.4", "37.7", "33", "48" ], [ "48.4", "49.5", "20.9", "40.1", "33.2", "49.1" ] ], "metrics_loc": "column", "metrics_type": "['P@1', 'P@1', 'P@1', 'P@1', 'P@1', 'P@1']", "target_entity": "['Without refinement', 'With refinement']", "table_html": "
Supervised || VecMap Supervised || Procrustes Without refinement || MUSE Without refinement || MUSEMethod With refinement || Our With refinement || Method
Similar language pairs || de-en 72.5 72 55.3 66.3 72.2 72.4
Similar language pairs || en-de 72.5 72.1 59.2 68.2 72.5 73.4
Similar language pairs || es-en 83.4 82.9 78.1 78 83.1 83.6
Similar language pairs || en-es 81.9 81.5 75.5 74.9 81.1 80.9
Similar language pairs || fr-en 81.5 82.2 72.2 75.6 81.2 83.3
Similar language pairs || en-fr 81.3 81.3 77.8 78.3 82.2 82.2
Similar language pairs || it-en 77.7 79.1 61 63.3 76.4 76.2
Similar language pairs || en-it 77.2 77.5 63.3 64.3 77.4 78.2
Metric || avg. 78.5 78.6 67.8 71.1 78.3 78.8
Distant language pairs || fi-en 55.9 56.2 0 42.8 28.6 53.5
Distant language pairs || en-fi 39.5 39.3 0 35.2 21 43
Distant language pairs || ru-en 60.5 61.3 43.8 54.3 49.9 58.9
Distant language pairs || en-ru 48 50.9 28.7 43.2 37.6 51.1
Distant language pairs || tr-en 55.7 58 12.2 42.1 25.5 54.7
Distant language pairs || en-tr 37.3 37.1 26.2 33.1 39.4 41.4
Distant language pairs || zh-en 45 41.2 26.9 32.1 30.8 42
Distant language pairs || en-zh 45.4 52.1 29.4 37.7 33 48
Metric || avg. 48.4 49.5 20.9 40.1 33.2 49.1
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D19-1450", "page_no": 6, "dir": "emnlp2019", "description": "Comparison with unsupervised GANs . As we have mentioned before, the preliminary mappings trained by the method of Lample et al. (2018) perform well for some similar language pairs, such as Spanish to English and French to English. After refinement, their unsupervised GANs models can reach the same level as supervised models for these similar language pairs. However, the biggest drawback of standard GANs is that they exhibit poor performance for distant language pairs. The results from Table 2 clearly confirm this. For example, without refinement, the mapping trained by the unsupervised GAN method can only correctly predict 12% of the words from Turkish to English. Given that the quality of preliminary mappings can seriously affect the effect of refinement, the low-quality preliminary mappings for distant language pairs severely limits the improvements brought by post-refinement. Notice that the method of Lample et al. (2018) scores a null result for English to Finnish on both BLI-1 and BLI2, indicating that totally unsupervised adversarial training can yields rather unpredictable results.", "class_sentence": "[2, 2, 2, 2, 1, 1, 1, 1]", "sentences": "['Comparison with unsupervised GANs .', 'As we have mentioned before, the preliminary mappings trained by the method of Lample et al. (2018) perform well for some similar language pairs, such as Spanish to English and French to English.', 'After refinement, their unsupervised GANs models can reach the same level as supervised models for these similar language pairs.', 'However, the biggest drawback of standard GANs is that they exhibit poor performance for distant language pairs.', 'The results from Table 2 clearly confirm this.', 'For example, without refinement, the mapping trained by the unsupervised GAN method can only correctly predict 12% of the words from Turkish to English.', 'Given that the quality of preliminary mappings can seriously affect the effect of refinement, the low-quality preliminary mappings for distant language pairs severely limits the improvements brought by post-refinement.', 'Notice that the method of Lample et al. (2018) scores a null result for English to Finnish on both BLI-1 and BLI2, indicating that totally unsupervised adversarial training can yields rather unpredictable results.']", "header_mention": "[None, None, None, None, None, ['Without refinement', 'MUSE', 'tr-en'], ['Without refinement', 'With refinement'], None]", "valid": 1, "image_id": "D19-1450-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results of bilingual lexicon induction (accuracy % P@1) for similar and distant language pairs on the dataset BLI-1. Procrustes and MUSE represent the supervised and unsupervised model of Lample et al. (2018), VecMap is the supervised model of Artetxe et al. (2017). Word translations are retrieved by using CSLS. Bold face indicates the best result overall and italics indicate the best result between the two columns without refinement.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Supervised || VecMapSupervised || ProcrustesWithout refinement || MUSEWithout refinement || MUSEMethodWith refinement || OurWith refinement || Method
Similar language pairs || de-en72.57255.366.372.272.4
Similar language pairs || en-de72.572.159.268.272.573.4
Similar language pairs || es-en83.482.978.17883.183.6
Similar language pairs || en-es81.981.575.574.981.180.9
Similar language pairs || fr-en81.582.272.275.681.283.3
Similar language pairs || en-fr81.381.377.878.382.282.2
Similar language pairs || it-en77.779.16163.376.476.2
Similar language pairs || en-it77.277.563.364.377.478.2
Metric || avg.78.578.667.871.178.378.8
Distant language pairs || fi-en55.956.2042.828.653.5
Distant language pairs || en-fi39.539.3035.22143
Distant language pairs || ru-en60.561.343.854.349.958.9
Distant language pairs || en-ru4850.928.743.237.651.1
Distant language pairs || tr-en55.75812.242.125.554.7
Distant language pairs || en-tr37.337.126.233.139.441.4
Distant language pairs || zh-en4541.226.932.130.842
Distant language pairs || en-zh45.452.129.437.73348
Metric || avg.48.449.520.940.133.249.1
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results of bilingual lexicon induction (accuracy \\% P@1) for similar and distant language pairs on the dataset BLI-1. Procrustes and MUSE represent the supervised and unsupervised model of Lample et al. (2018), VecMap is the supervised model of Artetxe et al. (2017). Word translations are retrieved by using CSLS. Bold face indicates the best result overall and italics indicate the best result between the two columns without refinement.}\n\\begin{tabular}{lrrrrrr}\n\\toprule\n & Supervised $\\parallel$ VecMap & Supervised $\\parallel$ Procrustes & Without refinement $\\parallel$ MUSE & Without refinement $\\parallel$ MUSEMethod & With refinement $\\parallel$ Our & With refinement $\\parallel$ Method \\\\\n\\midrule\nSimilar language pairs $\\parallel$ de-en & 72.5 & 72 & 55.3 & 66.3 & 72.2 & 72.4 \\\\\nSimilar language pairs $\\parallel$ en-de & 72.5 & 72.1 & 59.2 & 68.2 & 72.5 & 73.4 \\\\\nSimilar language pairs $\\parallel$ es-en & 83.4 & 82.9 & 78.1 & 78 & 83.1 & 83.6 \\\\\nSimilar language pairs $\\parallel$ en-es & 81.9 & 81.5 & 75.5 & 74.9 & 81.1 & 80.9 \\\\\nSimilar language pairs $\\parallel$ fr-en & 81.5 & 82.2 & 72.2 & 75.6 & 81.2 & 83.3 \\\\\nSimilar language pairs $\\parallel$ en-fr & 81.3 & 81.3 & 77.8 & 78.3 & 82.2 & 82.2 \\\\\nSimilar language pairs $\\parallel$ it-en & 77.7 & 79.1 & 61 & 63.3 & 76.4 & 76.2 \\\\\nSimilar language pairs $\\parallel$ en-it & 77.2 & 77.5 & 63.3 & 64.3 & 77.4 & 78.2 \\\\\nMetric $\\parallel$ avg. & 78.5 & 78.6 & 67.8 & 71.1 & 78.3 & 78.8 \\\\\nDistant language pairs $\\parallel$ fi-en & 55.9 & 56.2 & 0 & 42.8 & 28.6 & 53.5 \\\\\nDistant language pairs $\\parallel$ en-fi & 39.5 & 39.3 & 0 & 35.2 & 21 & 43 \\\\\nDistant language pairs $\\parallel$ ru-en & 60.5 & 61.3 & 43.8 & 54.3 & 49.9 & 58.9 \\\\\nDistant language pairs $\\parallel$ en-ru & 48 & 50.9 & 28.7 & 43.2 & 37.6 & 51.1 \\\\\nDistant language pairs $\\parallel$ tr-en & 55.7 & 58 & 12.2 & 42.1 & 25.5 & 54.7 \\\\\nDistant language pairs $\\parallel$ en-tr & 37.3 & 37.1 & 26.2 & 33.1 & 39.4 & 41.4 \\\\\nDistant language pairs $\\parallel$ zh-en & 45 & 41.2 & 26.9 & 32.1 & 30.8 & 42 \\\\\nDistant language pairs $\\parallel$ en-zh & 45.4 & 52.1 & 29.4 & 37.7 & 33 & 48 \\\\\nMetric $\\parallel$ avg. & 48.4 & 49.5 & 20.9 & 40.1 & 33.2 & 49.1 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Results of bilingual lexicon induction (accuracy % P@1) for similar and distant language pairs on the dataset BLI-1. Procrustes and MUSE represent the supervised and unsupervised model of Lample et al. (2018), VecMap is the supervised model of Artetxe et al. (2017). Word translations are retrieved by using CSLS. Bold face indicates the best result overall and italics indicate the best result between the two columns without refinement.\n
\n \n Supervised || VecMap\n \n Supervised || Procrustes\n \n Without refinement || MUSE\n \n Without refinement || MUSEMethod\n \n With refinement || Our\n \n With refinement || Method\n
\n Similar language pairs || de-en\n \n 72.5\n \n 72\n \n 55.3\n \n 66.3\n \n 72.2\n \n 72.4\n
\n Similar language pairs || en-de\n \n 72.5\n \n 72.1\n \n 59.2\n \n 68.2\n \n 72.5\n \n 73.4\n
\n Similar language pairs || es-en\n \n 83.4\n \n 82.9\n \n 78.1\n \n 78\n \n 83.1\n \n 83.6\n
\n Similar language pairs || en-es\n \n 81.9\n \n 81.5\n \n 75.5\n \n 74.9\n \n 81.1\n \n 80.9\n
\n Similar language pairs || fr-en\n \n 81.5\n \n 82.2\n \n 72.2\n \n 75.6\n \n 81.2\n \n 83.3\n
\n Similar language pairs || en-fr\n \n 81.3\n \n 81.3\n \n 77.8\n \n 78.3\n \n 82.2\n \n 82.2\n
\n Similar language pairs || it-en\n \n 77.7\n \n 79.1\n \n 61\n \n 63.3\n \n 76.4\n \n 76.2\n
\n Similar language pairs || en-it\n \n 77.2\n \n 77.5\n \n 63.3\n \n 64.3\n \n 77.4\n \n 78.2\n
\n Metric || avg.\n \n 78.5\n \n 78.6\n \n 67.8\n \n 71.1\n \n 78.3\n \n 78.8\n
\n Distant language pairs || fi-en\n \n 55.9\n \n 56.2\n \n 0\n \n 42.8\n \n 28.6\n \n 53.5\n
\n Distant language pairs || en-fi\n \n 39.5\n \n 39.3\n \n 0\n \n 35.2\n \n 21\n \n 43\n
\n Distant language pairs || ru-en\n \n 60.5\n \n 61.3\n \n 43.8\n \n 54.3\n \n 49.9\n \n 58.9\n
\n Distant language pairs || en-ru\n \n 48\n \n 50.9\n \n 28.7\n \n 43.2\n \n 37.6\n \n 51.1\n
\n Distant language pairs || tr-en\n \n 55.7\n \n 58\n \n 12.2\n \n 42.1\n \n 25.5\n \n 54.7\n
\n Distant language pairs || en-tr\n \n 37.3\n \n 37.1\n \n 26.2\n \n 33.1\n \n 39.4\n \n 41.4\n
\n Distant language pairs || zh-en\n \n 45\n \n 41.2\n \n 26.9\n \n 32.1\n \n 30.8\n \n 42\n
\n Distant language pairs || en-zh\n \n 45.4\n \n 52.1\n \n 29.4\n \n 37.7\n \n 33\n \n 48\n
\n Metric || avg.\n \n 48.4\n \n 49.5\n \n 20.9\n \n 40.1\n \n 33.2\n \n 49.1\n
\n" }, { "instance_id": "b9d2ec06-e31c-4a26-bdef-edb0a288fc87", "table_id_paper": "D19-1491table_1", "caption": "SEQ vs. CAMB system results on CWI", "row_header_level": 2, "row_headers": [ [ "Test Set", "NEWS" ], [ "Test Set", "WIKINEWS" ], [ "Test Set", "WIKIPEDIA" ] ], "column_header_level": 2, "column_headers": [ [ "Macro", "CAMB" ], [ "F-Score", "SEQ" ] ], "contents": [ [ "0.8633", "0.8763" ], [ "0.8317", "0.854" ], [ "0.778", "0.814" ] ], "metrics_loc": "column", "metrics_type": "['Macro', 'F-Score']", "target_entity": "['SEQ']", "table_html": "
Macro || CAMB F-Score || SEQ
Test Set || NEWS 0.8633 0.8763
Test Set || WIKINEWS 0.8317 0.854
Test Set || WIKIPEDIA 0.778 0.814
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "D19-1491", "page_no": 4, "dir": "emnlp2019", "description": "SEQ uses 300-dimensional GloVe embeddings as word representations (Pennington et al., 2014). The model is trained to predict the binary complexity of words as annotated in the dataset of Yimam et al. (2017). Training is performed over 20 iterations on randomly shuffled sentences from all genres included within the dataset. To test this novel architecture on CWI, we apply it to the CWI 2018 shared task test data (Yimam et al., 2018) and compare the results to the current state-of-theart (SOTA) CAMB system. Table 1 shows that the SEQ model outperforms the current SOTA system on all three text genres for binary CWI (statistically significant using McNemar's test, p=0.0016, chi-square=9.95).", "class_sentence": "[2, 2, 2, 2, 1]", "sentences": "['SEQ uses 300-dimensional GloVe embeddings as word representations (Pennington et al., 2014).', 'The model is trained to predict the binary complexity of words as annotated in the dataset of Yimam et al. (2017).', 'Training is performed over 20 iterations on randomly shuffled sentences from all genres included within the dataset.', 'To test this novel architecture on CWI, we apply it to the CWI 2018 shared task test data (Yimam et al., 2018) and compare the results to the current state-of-theart (SOTA) CAMB system.', \"Table 1 shows that the SEQ model outperforms the current SOTA system on all three text genres for binary CWI (statistically significant using McNemar's test, p=0.0016, chi-square=9.95).\"]", "header_mention": "[None, None, None, None, ['SEQ', 'NEWS', 'WIKINEWS', 'WIKIPEDIA']]", "valid": 1, "image_id": "D19-1491-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

SEQ vs. CAMB system results on CWI

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Macro || CAMBF-Score || SEQ
Test Set || NEWS0.86330.8763
Test Set || WIKINEWS0.83170.854
Test Set || WIKIPEDIA0.7780.814
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{SEQ vs. CAMB system results on CWI}\n\\begin{tabular}{lrr}\n\\toprule\n & Macro $\\parallel$ CAMB & F-Score $\\parallel$ SEQ \\\\\n\\midrule\nTest Set $\\parallel$ NEWS & 0.8633 & 0.8763 \\\\\nTest Set $\\parallel$ WIKINEWS & 0.8317 & 0.854 \\\\\nTest Set $\\parallel$ WIKIPEDIA & 0.778 & 0.814 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: SEQ vs. CAMB system results on CWI\n
\n \n Macro || CAMB\n \n F-Score || SEQ\n
\n Test Set || NEWS\n \n 0.8633\n \n 0.8763\n
\n Test Set || WIKINEWS\n \n 0.8317\n \n 0.854\n
\n Test Set || WIKIPEDIA\n \n 0.778\n \n 0.814\n
\n" }, { "instance_id": "0dd530e6-1b6b-49f7-9147-33e289443b1e", "table_id_paper": "D19-1497table_3", "caption": "Performance comparisons of different methods for citation count prediction using two datasets. \u201c\u2191\" ( \u201c\u2193\") indicates that a larger (smaller) value corresponds to a better performance.", "row_header_level": 3, "row_headers": [ [ "Datasets", "NIPS", "LR" ], [ "Datasets", "NIPS", "KNN" ], [ "Datasets", "NIPS", "SVR" ], [ "Datasets", "NIPS", "GBRT" ], [ "Datasets", "NIPS", "Wide&Deep" ], [ "Datasets", "NIPS", "MILAM" ], [ "Datasets", "NIPS", "Our model" ], [ "Datasets", "ICLR", "LR" ], [ "Datasets", "ICLR", "KNN" ], [ "Datasets", "ICLR", "SVR" ], [ "Datasets", "ICLR", "GBRT" ], [ "Datasets", "ICLR", "Wide&Deep" ], [ "Datasets", "ICLR", "MILAM" ], [ "Datasets", "ICLR", "Our model" ] ], "column_header_level": 1, "column_headers": [ [ "MAE" ], [ "RMSE" ], [ "OR(@30)" ], [ "OR(@50)" ], [ "Spearman Rank" ] ], "contents": [ [ "0.1776", "0.1903", "0.27", "0.33", "0.4776" ], [ "0.1701", "0.19", "0.33", "0.36", "0.4848" ], [ "0.1677", "0.1856", "0.33", "0.4", "0.5279" ], [ "0.1863", "0.1974", "0.23", "0.34", "0.531" ], [ "0.147", "0.1848", "0.3", "0.38", "0.5351" ], [ "0.1426", "0.1792", "0.37", "0.38", "0.5458" ], [ "0.1349", "0.1726", "0.4", "0.42", "0.5561" ], [ "0.2395", "0.2723", "0.4", "0.7", "0.1475" ], [ "0.2293", "0.2674", "0.4", "0.72", "0.1874" ], [ "0.2226", "0.2578", "0.4", "0.7", "0.1328" ], [ "0.2223", "0.2607", "0.43", "0.7", "0.1469" ], [ "0.2182", "0.2607", "0.47", "0.72", "0.244" ], [ "0.2093", "0.251", "0.47", "0.72", "0.251" ], [ "0.1866", "0.2279", "0.5", "0.76", "0.3026" ] ], "metrics_loc": "column", "metrics_type": "['MAE', 'RMSE', 'OR(@30)', 'OR(@50)', 'Spearman Rank']", "target_entity": "['Our model']", "table_html": "
MAE RMSE OR(@30) OR(@50) Spearman Rank
Datasets || NIPS || LR 0.1776 0.1903 0.27 0.33 0.4776
Datasets || NIPS || KNN 0.1701 0.19 0.33 0.36 0.4848
Datasets || NIPS || SVR 0.1677 0.1856 0.33 0.4 0.5279
Datasets || NIPS || GBRT 0.1863 0.1974 0.23 0.34 0.531
Datasets || NIPS || Wide&Deep 0.147 0.1848 0.3 0.38 0.5351
Datasets || NIPS || MILAM 0.1426 0.1792 0.37 0.38 0.5458
Datasets || NIPS || Our model 0.1349 0.1726 0.4 0.42 0.5561
Datasets || ICLR || LR 0.2395 0.2723 0.4 0.7 0.1475
Datasets || ICLR || KNN 0.2293 0.2674 0.4 0.72 0.1874
Datasets || ICLR || SVR 0.2226 0.2578 0.4 0.7 0.1328
Datasets || ICLR || GBRT 0.2223 0.2607 0.43 0.7 0.1469
Datasets || ICLR || Wide&Deep 0.2182 0.2607 0.47 0.72 0.244
Datasets || ICLR || MILAM 0.2093 0.251 0.47 0.72 0.251
Datasets || ICLR || Our model 0.1866 0.2279 0.5 0.76 0.3026
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "D19-1497", "page_no": 7, "dir": "emnlp2019", "description": "Main Results. Table 3 presents the performance of different methods on citation count prediction. We can make the following observations. First, the four traditional baselines (LR, KNN, SVR and GBRT) perform worse than the two deep learning baselines (W&D, MILAM). These four baselines only utilize the wide features with traditional machine learning models. Second, MILAM performs consistently better than W&D, since it has designed a more elaborate architecture to model the review text. Finally, our model outperforms all the baselines with a substantial margin, especially for the ICLR dataset. Our model is able to integrate the wide features and learn the comprehensive representation for peer review text, which is the key of the performance the two improvement over baselines. Overall, datasets show the similar findings. In what follows, we will report the results on ICLR dataset due to space limit.", "class_sentence": "[2, 1, 1, 1, 2, 1, 1, 2, 2, 2]", "sentences": "['Main Results.', 'Table 3 presents the performance of different methods on citation count prediction.', 'We can make the following observations.', 'First, the four traditional baselines (LR, KNN, SVR and GBRT) perform worse than the two deep learning baselines (W&D, MILAM).', 'These four baselines only utilize the wide features with traditional machine learning models.', 'Second, MILAM performs consistently better than W&D, since it has designed a more elaborate architecture to model the review text.', 'Finally, our model outperforms all the baselines with a substantial margin, especially for the ICLR dataset.', 'Our model is able to integrate the wide features and learn the comprehensive representation for peer review text, which is the key of the performance the two improvement over baselines.', 'Overall, datasets show the similar findings.', 'In what follows, we will report the results on ICLR dataset due to space limit.']", "header_mention": "[None, None, None, ['LR', 'KNN', 'SVR', 'GBRT', 'MILAM'], ['LR', 'KNN', 'SVR', 'GBRT'], ['MILAM'], ['Our model', 'ICLR'], ['Our model'], None, None]", "valid": 1, "image_id": "D19-1497-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance comparisons of different methods for citation count prediction using two datasets. \u201c\u2191" ( \u201c\u2193") indicates that a larger (smaller) value corresponds to a better performance.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n MAERMSEOR(@30)OR(@50)Spearman Rank
Datasets || NIPS || LR0.17760.19030.270.330.4776
Datasets || NIPS || KNN0.17010.190.330.360.4848
Datasets || NIPS || SVR0.16770.18560.330.40.5279
Datasets || NIPS || GBRT0.18630.19740.230.340.531
Datasets || NIPS || Wide&Deep0.1470.18480.30.380.5351
Datasets || NIPS || MILAM0.14260.17920.370.380.5458
Datasets || NIPS || Our model0.13490.17260.40.420.5561
Datasets || ICLR || LR0.23950.27230.40.70.1475
Datasets || ICLR || KNN0.22930.26740.40.720.1874
Datasets || ICLR || SVR0.22260.25780.40.70.1328
Datasets || ICLR || GBRT0.22230.26070.430.70.1469
Datasets || ICLR || Wide&Deep0.21820.26070.470.720.244
Datasets || ICLR || MILAM0.20930.2510.470.720.251
Datasets || ICLR || Our model0.18660.22790.50.760.3026
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance comparisons of different methods for citation count prediction using two datasets. \u201c$\\uparrow$\" ( \u201c$\\downarrow$\") indicates that a larger (smaller) value corresponds to a better performance.}\n\\begin{tabular}{lrrrrr}\n\\toprule\n & MAE & RMSE & OR(@30) & OR(@50) & Spearman Rank \\\\\n\\midrule\nDatasets $\\parallel$ NIPS $\\parallel$ LR & 0.1776 & 0.1903 & 0.27 & 0.33 & 0.4776 \\\\\nDatasets $\\parallel$ NIPS $\\parallel$ KNN & 0.1701 & 0.19 & 0.33 & 0.36 & 0.4848 \\\\\nDatasets $\\parallel$ NIPS $\\parallel$ SVR & 0.1677 & 0.1856 & 0.33 & 0.4 & 0.5279 \\\\\nDatasets $\\parallel$ NIPS $\\parallel$ GBRT & 0.1863 & 0.1974 & 0.23 & 0.34 & 0.531 \\\\\nDatasets $\\parallel$ NIPS $\\parallel$ Wide\\&Deep & 0.147 & 0.1848 & 0.3 & 0.38 & 0.5351 \\\\\nDatasets $\\parallel$ NIPS $\\parallel$ MILAM & 0.1426 & 0.1792 & 0.37 & 0.38 & 0.5458 \\\\\nDatasets $\\parallel$ NIPS $\\parallel$ Our model & 0.1349 & 0.1726 & 0.4 & 0.42 & 0.5561 \\\\\nDatasets $\\parallel$ ICLR $\\parallel$ LR & 0.2395 & 0.2723 & 0.4 & 0.7 & 0.1475 \\\\\nDatasets $\\parallel$ ICLR $\\parallel$ KNN & 0.2293 & 0.2674 & 0.4 & 0.72 & 0.1874 \\\\\nDatasets $\\parallel$ ICLR $\\parallel$ SVR & 0.2226 & 0.2578 & 0.4 & 0.7 & 0.1328 \\\\\nDatasets $\\parallel$ ICLR $\\parallel$ GBRT & 0.2223 & 0.2607 & 0.43 & 0.7 & 0.1469 \\\\\nDatasets $\\parallel$ ICLR $\\parallel$ Wide\\&Deep & 0.2182 & 0.2607 & 0.47 & 0.72 & 0.244 \\\\\nDatasets $\\parallel$ ICLR $\\parallel$ MILAM & 0.2093 & 0.251 & 0.47 & 0.72 & 0.251 \\\\\nDatasets $\\parallel$ ICLR $\\parallel$ Our model & 0.1866 & 0.2279 & 0.5 & 0.76 & 0.3026 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Performance comparisons of different methods for citation count prediction using two datasets. \u201c\u2191\" ( \u201c\u2193\") indicates that a larger (smaller) value corresponds to a better performance.\n
\n \n MAE\n \n RMSE\n \n OR(@30)\n \n OR(@50)\n \n Spearman Rank\n
\n Datasets || NIPS || LR\n \n 0.1776\n \n 0.1903\n \n 0.27\n \n 0.33\n \n 0.4776\n
\n Datasets || NIPS || KNN\n \n 0.1701\n \n 0.19\n \n 0.33\n \n 0.36\n \n 0.4848\n
\n Datasets || NIPS || SVR\n \n 0.1677\n \n 0.1856\n \n 0.33\n \n 0.4\n \n 0.5279\n
\n Datasets || NIPS || GBRT\n \n 0.1863\n \n 0.1974\n \n 0.23\n \n 0.34\n \n 0.531\n
\n Datasets || NIPS || Wide&Deep\n \n 0.147\n \n 0.1848\n \n 0.3\n \n 0.38\n \n 0.5351\n
\n Datasets || NIPS || MILAM\n \n 0.1426\n \n 0.1792\n \n 0.37\n \n 0.38\n \n 0.5458\n
\n Datasets || NIPS || Our model\n \n 0.1349\n \n 0.1726\n \n 0.4\n \n 0.42\n \n 0.5561\n
\n Datasets || ICLR || LR\n \n 0.2395\n \n 0.2723\n \n 0.4\n \n 0.7\n \n 0.1475\n
\n Datasets || ICLR || KNN\n \n 0.2293\n \n 0.2674\n \n 0.4\n \n 0.72\n \n 0.1874\n
\n Datasets || ICLR || SVR\n \n 0.2226\n \n 0.2578\n \n 0.4\n \n 0.7\n \n 0.1328\n
\n Datasets || ICLR || GBRT\n \n 0.2223\n \n 0.2607\n \n 0.43\n \n 0.7\n \n 0.1469\n
\n Datasets || ICLR || Wide&Deep\n \n 0.2182\n \n 0.2607\n \n 0.47\n \n 0.72\n \n 0.244\n
\n Datasets || ICLR || MILAM\n \n 0.2093\n \n 0.251\n \n 0.47\n \n 0.72\n \n 0.251\n
\n Datasets || ICLR || Our model\n \n 0.1866\n \n 0.2279\n \n 0.5\n \n 0.76\n \n 0.3026\n
\n" }, { "instance_id": "88da0199-60a7-4746-92b1-eca0034e9c0c", "table_id_paper": "D19-1507table_1", "caption": "Results of completion generation. We group MPC, character language model baseline, and two subword language models separately. +R implies the retrace algorithm. +M implies reranking with approximate marginalization. QPS stands for query per seconds. The higher the QPS, the better. The best results for each column related to accuracy are shown in bold for each segmentation algorithm (BPE and SR). SR model shows higher unseen PMRR scores (underlined). Our models are faster than the character baseline.", "row_header_level": 2, "row_headers": [ [ "Model", "MPC" ], [ "Model", "Char" ], [ "Model", "BPE" ], [ "Model", "BPE+R1" ], [ "Model", "BPE+R2" ], [ "Model", "BPE+R\u221e" ], [ "Model", "SR" ], [ "Model", "SR+M" ], [ "Model", "SR+R\u221e" ], [ "Model", "SR+R\u221e+M" ] ], "column_header_level": 1, "column_headers": [ [ "MRR Seen" ], [ "MRR Unseen" ], [ "MRR All" ], [ "PMRR Seen" ], [ "PMRR Unseen" ], [ "PMRR All" ], [ "MRL Seen" ], [ "MRL Unseen" ], [ "MRL All" ], [ "Execution Speed (QPS) CPU" ], [ "Execution Speed (QPS) GPU" ], [ "Decode Length" ] ], "contents": [ [ ".570", ".000", ".290", ".616", ".095", ".360", "8.06", "0.00", "4.10", ">100", ">100", "N/A" ], [ ".458", ".160", ".311", ".552", ".372", ".464", "5.77", "4.24", "5.02", "11.0 (1.0x)", "16.5 (1.0x)", "14.5" ], [ ".242", ".085", ".164", ".305", ".232", ".269", "0.49", "0.54", "0.51", "24.2 (2.2x)", "37.4 (2.3x)", "7.1" ], [ ".427", ".156", ".294", ".517", ".368", ".444", "5.28", "3.98", "4.64", "15.8 (1.4x)", "27.3 (1.7x)", "11.8" ], [ ".430", ".157", ".296", ".520", ".369", ".446", "5.44", "4.01", "4.74", "15.5 (1.4x)", "27.2 (1.6x)", "12.2" ], [ ".431", ".157", ".296", ".520", ".369", ".446", "5.50", "4.01", "4.76", "15.3 (1.4x)", "26.9 (1.6x)", "12.2" ], [ ".422", ".148", ".288", ".541", ".379", ".461", "5.11", "3.82", "4.48", "20.8 (1.9x)", "40.1 (2.4x)", "6.8" ], [ ".424", ".149", ".289", ".535", ".373", ".455", "5.14", "3.85", "4.50", "19.6 (1.8x)", "40.0 (2.4x)", "6.8" ], [ ".423", ".148", ".289", ".541", ".378", ".461", "5.14", "3.83", "4.50", "16.3 (1.5x)", "29.6 (1.8x)", "7.4" ], [ ".427", ".150", ".291", ".538", ".375", ".458", "5.19", "3.88", "4.54", "16.2 (1.5x)", "28.7 (1.7x)", "7.4" ] ], "metrics_loc": "column", "metrics_type": "['MRR Seen', 'MRR Unseen', 'MRR All', 'PMRR Seen', 'PMRR Unseen', 'PMRR All', 'MRL Seen', 'MRL Unseen', 'MRL All', 'Execution Speed (QPS) CPU', 'Execution Speed (QPS) GPU', 'Decode Length']", "target_entity": "['BPE', 'SR', 'BPE+R1', 'BPE+R2', 'BPE+R\u221e', 'SR+M', 'SR+R\u221e', 'SR+R\u221e+M']", "table_html": "
MRR Seen MRR Unseen MRR All PMRR Seen PMRR Unseen PMRR All MRL Seen MRL Unseen MRL All Execution Speed (QPS) CPU Execution Speed (QPS) GPU Decode Length
Model || MPC .570 .000 .290 .616 .095 .360 8.06 0.00 4.10 >100 >100 N/A
Model || Char .458 .160 .311 .552 .372 .464 5.77 4.24 5.02 11.0 (1.0x) 16.5 (1.0x) 14.5
Model || BPE .242 .085 .164 .305 .232 .269 0.49 0.54 0.51 24.2 (2.2x) 37.4 (2.3x) 7.1
Model || BPE+R1 .427 .156 .294 .517 .368 .444 5.28 3.98 4.64 15.8 (1.4x) 27.3 (1.7x) 11.8
Model || BPE+R2 .430 .157 .296 .520 .369 .446 5.44 4.01 4.74 15.5 (1.4x) 27.2 (1.6x) 12.2
Model || BPE+R\u221e .431 .157 .296 .520 .369 .446 5.50 4.01 4.76 15.3 (1.4x) 26.9 (1.6x) 12.2
Model || SR .422 .148 .288 .541 .379 .461 5.11 3.82 4.48 20.8 (1.9x) 40.1 (2.4x) 6.8
Model || SR+M .424 .149 .289 .535 .373 .455 5.14 3.85 4.50 19.6 (1.8x) 40.0 (2.4x) 6.8
Model || SR+R\u221e .423 .148 .289 .541 .378 .461 5.14 3.83 4.50 16.3 (1.5x) 29.6 (1.8x) 7.4
Model || SR+R\u221e+M .427 .150 .291 .538 .375 .458 5.19 3.88 4.54 16.2 (1.5x) 28.7 (1.7x) 7.4
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "D19-1507", "page_no": 6, "dir": "emnlp2019", "description": "6.3 Decoding Results . We performed comprehensive experiments to analyze the performance of query auto-completion. Table 1 shows the generation result of MPC, the character baseline, and our model variants. For BPE models, we varied the maximum retrace step to 0 (without retrace algorithm), 1, 2, and \u00e2\u02c6\u017e (no limitation on retracing step size). For SR models, we compare decoding results without any techniques, with marginalization only, with retrace algorithm only, and with both. MPC is a very fast and remarkably strong baseline. It is worse than language models in the overall score (MRR, PMRR, and MRL), but better for previously seen queries. However, it is unable to predict unseen queries. Even with efficient data structures, MPC requires huge memory to keep statistics of all previous queries. As a practical view, combining frequency-based traditional method and neural language model approach can boost the accuracy and meet trade-off between the performance and computational costs. MRRs and PMRRs of our best methods are close to that of the character model with less than 0.02 point drop. Notably, the SR model has better generalization ability in that their PMRR for unseen queries is higher than that of the character model. In a real scenario, it is more critical because unseen queries come in increasingly as time goes by.", "class_sentence": "[2, 1, 1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 2]", "sentences": "['6.3 Decoding Results .', 'We performed comprehensive experiments to analyze the performance of query auto-completion.', 'Table 1 shows the generation result of MPC, the character baseline, and our model variants.', 'For BPE models, we varied the maximum retrace step to 0 (without retrace algorithm), 1, 2, and \u00e2\u02c6\u017e (no limitation on retracing step size).', 'For SR models, we compare decoding results without any techniques, with marginalization only, with retrace algorithm only, and with both.', 'MPC is a very fast and remarkably strong baseline.', 'It is worse than language models in the overall score (MRR, PMRR, and MRL), but better for previously seen queries.', 'However, it is unable to predict unseen queries.', 'Even with efficient data structures, MPC requires huge memory to keep statistics of all previous queries.', 'As a practical view, combining frequency-based traditional method and neural language model approach can boost the accuracy and meet trade-off between the performance and computational costs.', 'MRRs and PMRRs of our best methods are close to that of the character model with less than 0.02 point drop.', 'Notably, the SR model has better generalization ability in that their PMRR for unseen queries is higher than that of the character model.', 'In a real scenario, it is more critical because unseen queries come in increasingly as time goes by.']", "header_mention": "[None, None, ['MPC', 'Char', 'BPE', 'SR'], ['BPE', 'BPE+R1', 'BPE+R2', 'BPE+R\u221e'], ['SR', 'SR+M', 'SR+R\u221e', 'SR+R\u221e+M'], ['MPC'], ['MPC', 'MRR Unseen', 'PMRR Unseen', 'MRL Unseen', 'MRR Seen', 'PMRR Seen', 'MRL Seen'], ['MPC', 'MRR Unseen', 'PMRR Unseen', 'MRL Unseen'], ['MPC'], None, ['MRR Seen', 'MRR Unseen', 'MRR All', 'PMRR Seen', 'PMRR Unseen', 'PMRR All', 'SR', 'BPE'], ['SR', 'PMRR Unseen', 'Char'], ['MRR Unseen', 'PMRR Unseen', 'MRL Unseen']]", "valid": 1, "image_id": "D19-1507-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results of completion generation. We group MPC, character language model baseline, and two subword language models separately. +R implies the retrace algorithm. +M implies reranking with approximate marginalization. QPS stands for query per seconds. The higher the QPS, the better. The best results for each column related to accuracy are shown in bold for each segmentation algorithm (BPE and SR). SR model shows higher unseen PMRR scores (underlined). Our models are faster than the character baseline.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n MRR SeenMRR UnseenMRR AllPMRR SeenPMRR UnseenPMRR AllMRL SeenMRL UnseenMRL AllExecution Speed (QPS) CPUExecution Speed (QPS) GPUDecode Length
Model || MPC.570.000.290.616.095.3608.060.004.10>100>100N/A
Model || Char.458.160.311.552.372.4645.774.245.0211.0 (1.0x)16.5 (1.0x)14.5
Model || BPE.242.085.164.305.232.2690.490.540.5124.2 (2.2x)37.4 (2.3x)7.1
Model || BPE+R1.427.156.294.517.368.4445.283.984.6415.8 (1.4x)27.3 (1.7x)11.8
Model || BPE+R2.430.157.296.520.369.4465.444.014.7415.5 (1.4x)27.2 (1.6x)12.2
Model || BPE+R\u221e.431.157.296.520.369.4465.504.014.7615.3 (1.4x)26.9 (1.6x)12.2
Model || SR.422.148.288.541.379.4615.113.824.4820.8 (1.9x)40.1 (2.4x)6.8
Model || SR+M.424.149.289.535.373.4555.143.854.5019.6 (1.8x)40.0 (2.4x)6.8
Model || SR+R\u221e.423.148.289.541.378.4615.143.834.5016.3 (1.5x)29.6 (1.8x)7.4
Model || SR+R\u221e+M.427.150.291.538.375.4585.193.884.5416.2 (1.5x)28.7 (1.7x)7.4
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results of completion generation. We group MPC, character language model baseline, and two subword language models separately. +R implies the retrace algorithm. +M implies reranking with approximate marginalization. QPS stands for query per seconds. The higher the QPS, the better. The best results for each column related to accuracy are shown in bold for each segmentation algorithm (BPE and SR). SR model shows higher unseen PMRR scores (underlined). Our models are faster than the character baseline.}\n\\begin{tabular}{lrrrrrrrrrlll}\n\\toprule\n & MRR Seen & MRR Unseen & MRR All & PMRR Seen & PMRR Unseen & PMRR All & MRL Seen & MRL Unseen & MRL All & Execution Speed (QPS) CPU & Execution Speed (QPS) GPU & Decode Length \\\\\n\\midrule\nModel $\\parallel$ MPC & 0.57 & 0 & 0.29 & 0.616 & 0.095 & 0.36 & 8.06 & 0 & 4.1 & $>$100 & $>$100 & \\\\\nModel $\\parallel$ Char & 0.458 & 0.16 & 0.311 & 0.552 & 0.372 & 0.464 & 5.77 & 4.24 & 5.02 & 11.0 (1.0x) & 16.5 (1.0x) & 14.5 \\\\\nModel $\\parallel$ BPE & 0.242 & 0.085 & 0.164 & 0.305 & 0.232 & 0.269 & 0.49 & 0.54 & 0.51 & 24.2 (2.2x) & 37.4 (2.3x) & 7.1 \\\\\nModel $\\parallel$ BPE+R1 & 0.427 & 0.156 & 0.294 & 0.517 & 0.368 & 0.444 & 5.28 & 3.98 & 4.64 & 15.8 (1.4x) & 27.3 (1.7x) & 11.8 \\\\\nModel $\\parallel$ BPE+R2 & 0.43 & 0.157 & 0.296 & 0.52 & 0.369 & 0.446 & 5.44 & 4.01 & 4.74 & 15.5 (1.4x) & 27.2 (1.6x) & 12.2 \\\\\nModel $\\parallel$ BPE+R$\\infty$ & 0.431 & 0.157 & 0.296 & 0.52 & 0.369 & 0.446 & 5.5 & 4.01 & 4.76 & 15.3 (1.4x) & 26.9 (1.6x) & 12.2 \\\\\nModel $\\parallel$ SR & 0.422 & 0.148 & 0.288 & 0.541 & 0.379 & 0.461 & 5.11 & 3.82 & 4.48 & 20.8 (1.9x) & 40.1 (2.4x) & 6.8 \\\\\nModel $\\parallel$ SR+M & 0.424 & 0.149 & 0.289 & 0.535 & 0.373 & 0.455 & 5.14 & 3.85 & 4.5 & 19.6 (1.8x) & 40.0 (2.4x) & 6.8 \\\\\nModel $\\parallel$ SR+R$\\infty$ & 0.423 & 0.148 & 0.289 & 0.541 & 0.378 & 0.461 & 5.14 & 3.83 & 4.5 & 16.3 (1.5x) & 29.6 (1.8x) & 7.4 \\\\\nModel $\\parallel$ SR+R$\\infty$+M & 0.427 & 0.15 & 0.291 & 0.538 & 0.375 & 0.458 & 5.19 & 3.88 & 4.54 & 16.2 (1.5x) & 28.7 (1.7x) & 7.4 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Results of completion generation. We group MPC, character language model baseline, and two subword language models separately. +R implies the retrace algorithm. +M implies reranking with approximate marginalization. QPS stands for query per seconds. The higher the QPS, the better. The best results for each column related to accuracy are shown in bold for each segmentation algorithm (BPE and SR). SR model shows higher unseen PMRR scores (underlined). Our models are faster than the character baseline.\n
\n \n MRR Seen\n \n MRR Unseen\n \n MRR All\n \n PMRR Seen\n \n PMRR Unseen\n \n PMRR All\n \n MRL Seen\n \n MRL Unseen\n \n MRL All\n \n Execution Speed (QPS) CPU\n \n Execution Speed (QPS) GPU\n \n Decode Length\n
\n Model || MPC\n \n .570\n \n .000\n \n .290\n \n .616\n \n .095\n \n .360\n \n 8.06\n \n 0.00\n \n 4.10\n \n >100\n \n >100\n \n N/A\n
\n Model || Char\n \n .458\n \n .160\n \n .311\n \n .552\n \n .372\n \n .464\n \n 5.77\n \n 4.24\n \n 5.02\n \n 11.0 (1.0x)\n \n 16.5 (1.0x)\n \n 14.5\n
\n Model || BPE\n \n .242\n \n .085\n \n .164\n \n .305\n \n .232\n \n .269\n \n 0.49\n \n 0.54\n \n 0.51\n \n 24.2 (2.2x)\n \n 37.4 (2.3x)\n \n 7.1\n
\n Model || BPE+R1\n \n .427\n \n .156\n \n .294\n \n .517\n \n .368\n \n .444\n \n 5.28\n \n 3.98\n \n 4.64\n \n 15.8 (1.4x)\n \n 27.3 (1.7x)\n \n 11.8\n
\n Model || BPE+R2\n \n .430\n \n .157\n \n .296\n \n .520\n \n .369\n \n .446\n \n 5.44\n \n 4.01\n \n 4.74\n \n 15.5 (1.4x)\n \n 27.2 (1.6x)\n \n 12.2\n
\n Model || BPE+R\u221e\n \n .431\n \n .157\n \n .296\n \n .520\n \n .369\n \n .446\n \n 5.50\n \n 4.01\n \n 4.76\n \n 15.3 (1.4x)\n \n 26.9 (1.6x)\n \n 12.2\n
\n Model || SR\n \n .422\n \n .148\n \n .288\n \n .541\n \n .379\n \n .461\n \n 5.11\n \n 3.82\n \n 4.48\n \n 20.8 (1.9x)\n \n 40.1 (2.4x)\n \n 6.8\n
\n Model || SR+M\n \n .424\n \n .149\n \n .289\n \n .535\n \n .373\n \n .455\n \n 5.14\n \n 3.85\n \n 4.50\n \n 19.6 (1.8x)\n \n 40.0 (2.4x)\n \n 6.8\n
\n Model || SR+R\u221e\n \n .423\n \n .148\n \n .289\n \n .541\n \n .378\n \n .461\n \n 5.14\n \n 3.83\n \n 4.50\n \n 16.3 (1.5x)\n \n 29.6 (1.8x)\n \n 7.4\n
\n Model || SR+R\u221e+M\n \n .427\n \n .150\n \n .291\n \n .538\n \n .375\n \n .458\n \n 5.19\n \n 3.88\n \n 4.54\n \n 16.2 (1.5x)\n \n 28.7 (1.7x)\n \n 7.4\n
\n" }, { "instance_id": "9930f7f6-a148-49b6-b52a-0b7295251864", "table_id_paper": "D19-1512table_4", "caption": "Evaluation results on automatic metrics and human judgment. Human evaluation results are calculated by combining labels from the three judges. \u201cKappa\u201d means Fleiss\u2019 kappa. Numbers in bold mean that improvement over the best baseline is statistically significant.", "row_header_level": 4, "row_headers": [ [ "Dataset", "Tencent", "Models", "IR-T" ], [ "Dataset", "Tencent", "Models", "IR-TC" ], [ "Dataset", "Tencent", "Models", "Seq2seq" ], [ "Dataset", "Tencent", "Models", "Att" ], [ "Dataset", "Tencent", "Models", "Att-TC" ], [ "Dataset", "Tencent", "Models", "GANN" ], [ "Dataset", "Tencent", "Models", "DeepCom" ], [ "Dataset", "Yahoo", "Models", "IR-T" ], [ "Dataset", "Yahoo", "Models", "IR-TC" ], [ "Dataset", "Yahoo", "Models", "Seq2seq" ], [ "Dataset", "Yahoo", "Models", "Att" ], [ "Dataset", "Yahoo", "Models", "Att-TC" ], [ "Dataset", "Yahoo", "Models", "GANN" ], [ "Dataset", "Yahoo", "Models", "DeepCom" ] ], "column_header_level": 1, "column_headers": [ [ "METEOR" ], [ "W-METEOR" ], [ "Rouge_L" ], [ "W-Rouge_L" ], [ "CIDEr" ], [ "W-CIDEr" ], [ "BLEU-1" ], [ "W-BLEU-1" ], [ "Human" ], [ "Kappa" ] ], "contents": [ [ "0.107", "0.086", "0.254", "0.217", "0.018", "0.014", "0.495", "0.470", "2.43", "0.64" ], [ "0.127", "0.101", "0.266", "0.225", "0.056", "0.044", "0.474", "0.436", "2.57", "0.71" ], [ "0.064", "0.047", "0.196", "0.150", "0.011", "0.008", "0.374", "0.320", "1.68", "0.83" ], [ "0.080", "0.058", "0.246", "0.186", "0.010", "0.007", "0.481", "0.453", "1.81", "0.79" ], [ "0.114", "0.082", "0.299", "0.223", "0.023", "0.017", "0.602", "0.551", "2.26", "0.69" ], [ "0.097", "0.075", "0.282", "0.222", "0.010", "0.008", "0.312", "0.278", "2.06", "0.73" ], [ "0.181", "0.138", "0.317", "0.250", "0.029", "0.023", "0.721", "0.656", "3.58", "0.65" ], [ "0.114", "-", "0.214", "-", "0.014", "-", "0.472", "-", "2.71", "0.67" ], [ "0.117", "-", "0.219", "-", "0.017", "-", "0.483", "-", "2.86", "0.61" ], [ "0.061", "-", "0.203", "-", "0.011", "-", "0.365", "-", "2.26", "0.68" ], [ "0.075", "-", "0.217", "-", "0.017", "-", "0.462", "-", "2.29", "0.78" ], [ "0.089", "-", "0.246", "-", "0.022", "-", "0.515", "-", "2.74", "0.63" ], [ "0.079", "-", "0.228", "-", "0.019", "-", "0.496", "-", "2.52", "0.64" ], [ "0.107", "-", "0.263", "-", "0.024", "-", "0.665", "-", "3.35", "0.68" ] ], "metrics_loc": "column", "metrics_type": "['METEOR', 'W-METEOR', 'Rouge_L', 'W-Rouge_L', 'CIDEr', 'W-CIDEr', 'BLEU-1', 'W-BLEU-1', 'Human', 'Kappa']", "target_entity": "['DeepCom']", "table_html": "
METEOR W-METEOR Rouge_L W-Rouge_L CIDEr W-CIDEr BLEU-1 W-BLEU-1 Human Kappa
Dataset || Tencent || Models || IR-T 0.107 0.086 0.254 0.217 0.018 0.014 0.495 0.470 2.43 0.64
Dataset || Tencent || Models || IR-TC 0.127 0.101 0.266 0.225 0.056 0.044 0.474 0.436 2.57 0.71
Dataset || Tencent || Models || Seq2seq 0.064 0.047 0.196 0.150 0.011 0.008 0.374 0.320 1.68 0.83
Dataset || Tencent || Models || Att 0.080 0.058 0.246 0.186 0.010 0.007 0.481 0.453 1.81 0.79
Dataset || Tencent || Models || Att-TC 0.114 0.082 0.299 0.223 0.023 0.017 0.602 0.551 2.26 0.69
Dataset || Tencent || Models || GANN 0.097 0.075 0.282 0.222 0.010 0.008 0.312 0.278 2.06 0.73
Dataset || Tencent || Models || DeepCom 0.181 0.138 0.317 0.250 0.029 0.023 0.721 0.656 3.58 0.65
Dataset || Yahoo || Models || IR-T 0.114 - 0.214 - 0.014 - 0.472 - 2.71 0.67
Dataset || Yahoo || Models || IR-TC 0.117 - 0.219 - 0.017 - 0.483 - 2.86 0.61
Dataset || Yahoo || Models || Seq2seq 0.061 - 0.203 - 0.011 - 0.365 - 2.26 0.68
Dataset || Yahoo || Models || Att 0.075 - 0.217 - 0.017 - 0.462 - 2.29 0.78
Dataset || Yahoo || Models || Att-TC 0.089 - 0.246 - 0.022 - 0.515 - 2.74 0.63
Dataset || Yahoo || Models || GANN 0.079 - 0.228 - 0.019 - 0.496 - 2.52 0.64
Dataset || Yahoo || Models || DeepCom 0.107 - 0.263 - 0.024 - 0.665 - 3.35 0.68
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D19-1512", "page_no": 8, "dir": "emnlp2019", "description": "4.4 Evaluation Results . Table 4 reports evaluation results in terms of both automatic metrics and human annotations. On most automatic metrics, DeepCom outperforms baseline methods, and the improvement is statistically significant (t-test with p-value < 0.01). The improvement on BLEU-1 and W-BLEU-1 is much bigger than that on other metrics. This is because BLEU-1 only measures the proportion of matched unigrams out of the total number of unigrams in the generated comments. In human evaluation, although the absolute numbers are different from those reported in (Qin et al., 2018) due to the difference between human judgements, the overall trend is consistent. In human evaluation, the values of Fleiss' kappa over all models are more than 0:6, indicating substantial agreement among the annotators. Although built in a complicated structure, GANN does not bring much improvement over other baseline methods, which demonstrates that only using news titles is not enough in comment generation. IR-TC and Att-TC represent the best retrieval model and the best generation model among the baselines on both datasets, implying that news bodies, even used in a simple way, can provide useful information to comment generation.", "class_sentence": "[0, 1, 2, 1, 2, 1, 1, 1, 1]", "sentences": "['4.4 Evaluation Results .', 'Table 4 reports evaluation results in terms of both automatic metrics and human annotations.', 'On most automatic metrics, DeepCom outperforms baseline methods, and the improvement is statistically significant (t-test with p-value < 0.01).', 'The improvement on BLEU-1 and W-BLEU-1 is much bigger than that on other metrics.', 'This is because BLEU-1 only measures the proportion of matched unigrams out of the total number of unigrams in the generated comments.', 'In human evaluation, although the absolute numbers are different from those reported in (Qin et al., 2018) due to the difference between human judgements, the overall trend is consistent.', \"In human evaluation, the values of Fleiss' kappa over all models are more than 0:6, indicating substantial agreement among the annotators.\", 'Although built in a complicated structure, GANN does not bring much improvement over other baseline methods, which demonstrates that only using news titles is not enough in comment generation.', 'IR-TC and Att-TC represent the best retrieval model and the best generation model among the baselines on both datasets, implying that news bodies, even used in a simple way, can provide useful information to comment generation.']", "header_mention": "[None, None, ['DeepCom'], ['DeepCom', 'BLEU-1', 'W-BLEU-1'], None, ['DeepCom', 'Human'], ['Models', 'Kappa'], ['GANN'], ['IR-TC', 'Att-TC', 'Dataset']]", "valid": 1, "image_id": "D19-1512-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Evaluation results on automatic metrics and human judgment. Human evaluation results are calculated by combining labels from the three judges. \u201cKappa\u201d means Fleiss\u2019 kappa. Numbers in bold mean that improvement over the best baseline is statistically significant.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n METEORW-METEORRouge_LW-Rouge_LCIDErW-CIDErBLEU-1W-BLEU-1HumanKappa
Dataset || Tencent || Models || IR-T0.1070.0860.2540.2170.0180.0140.4950.4702.430.64
Dataset || Tencent || Models || IR-TC0.1270.1010.2660.2250.0560.0440.4740.4362.570.71
Dataset || Tencent || Models || Seq2seq0.0640.0470.1960.1500.0110.0080.3740.3201.680.83
Dataset || Tencent || Models || Att0.0800.0580.2460.1860.0100.0070.4810.4531.810.79
Dataset || Tencent || Models || Att-TC0.1140.0820.2990.2230.0230.0170.6020.5512.260.69
Dataset || Tencent || Models || GANN0.0970.0750.2820.2220.0100.0080.3120.2782.060.73
Dataset || Tencent || Models || DeepCom0.1810.1380.3170.2500.0290.0230.7210.6563.580.65
Dataset || Yahoo || Models || IR-T0.114-0.214-0.014-0.472-2.710.67
Dataset || Yahoo || Models || IR-TC0.117-0.219-0.017-0.483-2.860.61
Dataset || Yahoo || Models || Seq2seq0.061-0.203-0.011-0.365-2.260.68
Dataset || Yahoo || Models || Att0.075-0.217-0.017-0.462-2.290.78
Dataset || Yahoo || Models || Att-TC0.089-0.246-0.022-0.515-2.740.63
Dataset || Yahoo || Models || GANN0.079-0.228-0.019-0.496-2.520.64
Dataset || Yahoo || Models || DeepCom0.107-0.263-0.024-0.665-3.350.68
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Evaluation results on automatic metrics and human judgment. Human evaluation results are calculated by combining labels from the three judges. \u201cKappa\u201d means Fleiss\u2019 kappa. Numbers in bold mean that improvement over the best baseline is statistically significant.}\n\\begin{tabular}{lrlrlrlrlrr}\n\\toprule\n & METEOR & W-METEOR & Rouge\\_L & W-Rouge\\_L & CIDEr & W-CIDEr & BLEU-1 & W-BLEU-1 & Human & Kappa \\\\\n\\midrule\nDataset $\\parallel$ Tencent $\\parallel$ Models $\\parallel$ IR-T & 0.107 & 0.086 & 0.254 & 0.217 & 0.018 & 0.014 & 0.495 & 0.470 & 2.43 & 0.64 \\\\\nDataset $\\parallel$ Tencent $\\parallel$ Models $\\parallel$ IR-TC & 0.127 & 0.101 & 0.266 & 0.225 & 0.056 & 0.044 & 0.474 & 0.436 & 2.57 & 0.71 \\\\\nDataset $\\parallel$ Tencent $\\parallel$ Models $\\parallel$ Seq2seq & 0.064 & 0.047 & 0.196 & 0.150 & 0.011 & 0.008 & 0.374 & 0.320 & 1.68 & 0.83 \\\\\nDataset $\\parallel$ Tencent $\\parallel$ Models $\\parallel$ Att & 0.08 & 0.058 & 0.246 & 0.186 & 0.01 & 0.007 & 0.481 & 0.453 & 1.81 & 0.79 \\\\\nDataset $\\parallel$ Tencent $\\parallel$ Models $\\parallel$ Att-TC & 0.114 & 0.082 & 0.299 & 0.223 & 0.023 & 0.017 & 0.602 & 0.551 & 2.26 & 0.69 \\\\\nDataset $\\parallel$ Tencent $\\parallel$ Models $\\parallel$ GANN & 0.097 & 0.075 & 0.282 & 0.222 & 0.01 & 0.008 & 0.312 & 0.278 & 2.06 & 0.73 \\\\\nDataset $\\parallel$ Tencent $\\parallel$ Models $\\parallel$ DeepCom & 0.181 & 0.138 & 0.317 & 0.250 & 0.029 & 0.023 & 0.721 & 0.656 & 3.58 & 0.65 \\\\\nDataset $\\parallel$ Yahoo $\\parallel$ Models $\\parallel$ IR-T & 0.114 & - & 0.214 & - & 0.014 & - & 0.472 & - & 2.71 & 0.67 \\\\\nDataset $\\parallel$ Yahoo $\\parallel$ Models $\\parallel$ IR-TC & 0.117 & - & 0.219 & - & 0.017 & - & 0.483 & - & 2.86 & 0.61 \\\\\nDataset $\\parallel$ Yahoo $\\parallel$ Models $\\parallel$ Seq2seq & 0.061 & - & 0.203 & - & 0.011 & - & 0.365 & - & 2.26 & 0.68 \\\\\nDataset $\\parallel$ Yahoo $\\parallel$ Models $\\parallel$ Att & 0.075 & - & 0.217 & - & 0.017 & - & 0.462 & - & 2.29 & 0.78 \\\\\nDataset $\\parallel$ Yahoo $\\parallel$ Models $\\parallel$ Att-TC & 0.089 & - & 0.246 & - & 0.022 & - & 0.515 & - & 2.74 & 0.63 \\\\\nDataset $\\parallel$ Yahoo $\\parallel$ Models $\\parallel$ GANN & 0.079 & - & 0.228 & - & 0.019 & - & 0.496 & - & 2.52 & 0.64 \\\\\nDataset $\\parallel$ Yahoo $\\parallel$ Models $\\parallel$ DeepCom & 0.107 & - & 0.263 & - & 0.024 & - & 0.665 & - & 3.35 & 0.68 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Evaluation results on automatic metrics and human judgment. Human evaluation results are calculated by combining labels from the three judges. \u201cKappa\u201d means Fleiss\u2019 kappa. Numbers in bold mean that improvement over the best baseline is statistically significant.\n
\n \n METEOR\n \n W-METEOR\n \n Rouge_L\n \n W-Rouge_L\n \n CIDEr\n \n W-CIDEr\n \n BLEU-1\n \n W-BLEU-1\n \n Human\n \n Kappa\n
\n Dataset || Tencent || Models || IR-T\n \n 0.107\n \n 0.086\n \n 0.254\n \n 0.217\n \n 0.018\n \n 0.014\n \n 0.495\n \n 0.470\n \n 2.43\n \n 0.64\n
\n Dataset || Tencent || Models || IR-TC\n \n 0.127\n \n 0.101\n \n 0.266\n \n 0.225\n \n 0.056\n \n 0.044\n \n 0.474\n \n 0.436\n \n 2.57\n \n 0.71\n
\n Dataset || Tencent || Models || Seq2seq\n \n 0.064\n \n 0.047\n \n 0.196\n \n 0.150\n \n 0.011\n \n 0.008\n \n 0.374\n \n 0.320\n \n 1.68\n \n 0.83\n
\n Dataset || Tencent || Models || Att\n \n 0.080\n \n 0.058\n \n 0.246\n \n 0.186\n \n 0.010\n \n 0.007\n \n 0.481\n \n 0.453\n \n 1.81\n \n 0.79\n
\n Dataset || Tencent || Models || Att-TC\n \n 0.114\n \n 0.082\n \n 0.299\n \n 0.223\n \n 0.023\n \n 0.017\n \n 0.602\n \n 0.551\n \n 2.26\n \n 0.69\n
\n Dataset || Tencent || Models || GANN\n \n 0.097\n \n 0.075\n \n 0.282\n \n 0.222\n \n 0.010\n \n 0.008\n \n 0.312\n \n 0.278\n \n 2.06\n \n 0.73\n
\n Dataset || Tencent || Models || DeepCom\n \n 0.181\n \n 0.138\n \n 0.317\n \n 0.250\n \n 0.029\n \n 0.023\n \n 0.721\n \n 0.656\n \n 3.58\n \n 0.65\n
\n Dataset || Yahoo || Models || IR-T\n \n 0.114\n \n -\n \n 0.214\n \n -\n \n 0.014\n \n -\n \n 0.472\n \n -\n \n 2.71\n \n 0.67\n
\n Dataset || Yahoo || Models || IR-TC\n \n 0.117\n \n -\n \n 0.219\n \n -\n \n 0.017\n \n -\n \n 0.483\n \n -\n \n 2.86\n \n 0.61\n
\n Dataset || Yahoo || Models || Seq2seq\n \n 0.061\n \n -\n \n 0.203\n \n -\n \n 0.011\n \n -\n \n 0.365\n \n -\n \n 2.26\n \n 0.68\n
\n Dataset || Yahoo || Models || Att\n \n 0.075\n \n -\n \n 0.217\n \n -\n \n 0.017\n \n -\n \n 0.462\n \n -\n \n 2.29\n \n 0.78\n
\n Dataset || Yahoo || Models || Att-TC\n \n 0.089\n \n -\n \n 0.246\n \n -\n \n 0.022\n \n -\n \n 0.515\n \n -\n \n 2.74\n \n 0.63\n
\n Dataset || Yahoo || Models || GANN\n \n 0.079\n \n -\n \n 0.228\n \n -\n \n 0.019\n \n -\n \n 0.496\n \n -\n \n 2.52\n \n 0.64\n
\n Dataset || Yahoo || Models || DeepCom\n \n 0.107\n \n -\n \n 0.263\n \n -\n \n 0.024\n \n -\n \n 0.665\n \n -\n \n 3.35\n \n 0.68\n
\n" }, { "instance_id": "915addbf-771d-4a13-b6d0-e52469bb8dc1", "table_id_paper": "D19-1581table_4", "caption": "Results for small labeled training data. Given the performance with the full dataset, we show BERT trained only with the AL data.", "row_header_level": 4, "row_headers": [ [ "Training dataset", "ACP (6K)", "Encoder", "BERT" ], [ "Training dataset", " +AL", "Encoder", "BERT" ], [ "Training dataset", "ACP (6K)", "Encoder", "BiGRU" ], [ "Training dataset", " +AL+CA+CO", "Encoder", "BiGRU" ] ], "column_header_level": 1, "column_headers": [ [ "Acc" ] ], "contents": [ [ "0.876" ], [ "0.886" ], [ "0.83" ], [ "0.879" ] ], "metrics_loc": "column", "metrics_type": "['Acc']", "target_entity": "['ACP (6K)']", "table_html": "
Acc
Training dataset || ACP (6K) || Encoder || BERT 0.876
Training dataset || +AL || Encoder || BERT 0.886
Training dataset || ACP (6K) || Encoder || BiGRU 0.83
Training dataset || +AL+CA+CO || Encoder || BiGRU 0.879
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D19-1581", "page_no": 6, "dir": "emnlp2019", "description": "Contrary to our expectations, supervised models (ACP) outperformed semi-supervised models (ACP+AL+CA+CO). This suggests that the training set of 0.6 million events is sufficiently large for training the models. For comparison, we trained the models with a subset (6,000 events) of the ACP dataset. As the results shown in Table 4 demonstrate, our method is effective when labeled data are small.", "class_sentence": "[1, 2, 2, 1]", "sentences": "['Contrary to our expectations, supervised models (ACP) outperformed semi-supervised models (ACP+AL+CA+CO).', 'This suggests that the training set of 0.6 million events is sufficiently large for training the models.', 'For comparison, we trained the models with a subset (6,000 events) of the ACP dataset.', 'As the results shown in Table 4 demonstrate, our method is effective when labeled data are small.']", "header_mention": "[['ACP (6K)', ' +AL+CA+CO'], None, ['ACP (6K)'], None]", "valid": 1, "image_id": "D19-1581-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results for small labeled training data. Given the performance with the full dataset, we show BERT trained only with the AL data.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Acc
Training dataset || ACP (6K) || Encoder || BERT0.876
Training dataset || +AL || Encoder || BERT0.886
Training dataset || ACP (6K) || Encoder || BiGRU0.83
Training dataset || +AL+CA+CO || Encoder || BiGRU0.879
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results for small labeled training data. Given the performance with the full dataset, we show BERT trained only with the AL data.}\n\\begin{tabular}{lr}\n\\toprule\n & Acc \\\\\n\\midrule\nTraining dataset $\\parallel$ ACP (6K) $\\parallel$ Encoder $\\parallel$ BERT & 0.876 \\\\\nTraining dataset $\\parallel$ +AL $\\parallel$ Encoder $\\parallel$ BERT & 0.886 \\\\\nTraining dataset $\\parallel$ ACP (6K) $\\parallel$ Encoder $\\parallel$ BiGRU & 0.83 \\\\\nTraining dataset $\\parallel$ +AL+CA+CO $\\parallel$ Encoder $\\parallel$ BiGRU & 0.879 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Results for small labeled training data. Given the performance with the full dataset, we show BERT trained only with the AL data.\n
\n \n Acc\n
\n Training dataset || ACP (6K) || Encoder || BERT\n \n 0.876\n
\n Training dataset || +AL || Encoder || BERT\n \n 0.886\n
\n Training dataset || ACP (6K) || Encoder || BiGRU\n \n 0.83\n
\n Training dataset || +AL+CA+CO || Encoder || BiGRU\n \n 0.879\n
\n" }, { "instance_id": "d1180799-af17-4762-8039-6f2a9b1d5f3e", "table_id_paper": "D19-1633table_4", "caption": "The performance (BLEU) of base CMLM with different amounts of mask-predict iterations (T ) on WMT\u201914 EN-DE, bucketed by target sequence length (N). Decoding with (cid:96) = 1 length candidates.", "row_header_level": 1, "row_headers": [ [ "1 ? N < 10" ], [ "10 ? N < 20" ], [ "20 ? N < 30" ], [ "30 ? N < 40" ], [ "40 ? N" ] ], "column_header_level": 1, "column_headers": [ [ "T = 4" ], [ "T = 10" ], [ "T = N" ] ], "contents": [ [ "21.8", "22.4", "22.4" ], [ "24.6", "25.9", "26" ], [ "24.9", "26.7", "27.1" ], [ "24.9", "26.7", "27.6" ], [ "25", "27.5", "28.1" ] ], "metrics_loc": "column", "metrics_type": "['BLEU', 'BLEU', 'BLEU']", "target_entity": "['T = N']", "table_html": "
T = 4 T = 10 T = N
1 ? N < 10 21.8 22.4 22.4
10 ? N < 20 24.6 25.9 26
20 ? N < 30 24.9 26.7 27.1
30 ? N < 40 24.9 26.7 27.6
40 ? N 25 27.5 28.1
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "D19-1633", "page_no": 7, "dir": "emnlp2019", "description": "Table 4 shows that increasing the number of decoding iterations (T) appears to mainly improve the performance on longer sequences. Having said that, the performance differences across length buckets are not very large, and it seems that even 4 mask-predict iterations are enough to produce decent translations for long sequences (40 ? N).", "class_sentence": "[1, 1]", "sentences": "['Table 4 shows that increasing the number of decoding iterations (T) appears to mainly improve the performance on longer sequences.', 'Having said that, the performance differences across length buckets are not very large, and it seems that even 4 mask-predict iterations are enough to produce decent translations for long sequences (40 ? N).']", "header_mention": "[['T = N'], None]", "valid": 1, "image_id": "D19-1633-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

The performance (BLEU) of base CMLM with different amounts of mask-predict iterations (T ) on WMT\u201914 EN-DE, bucketed by target sequence length (N). Decoding with (cid:96) = 1 length candidates.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n T = 4T = 10T = N
1 ? N < 1021.822.422.4
10 ? N < 2024.625.926
20 ? N < 3024.926.727.1
30 ? N < 4024.926.727.6
40 ? N2527.528.1
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{The performance (BLEU) of base CMLM with different amounts of mask-predict iterations (T ) on WMT\u201914 EN-DE, bucketed by target sequence length (N). Decoding with (cid:96) $=$ 1 length candidates.}\n\\begin{tabular}{lrrr}\n\\toprule\n & T $=$ 4 & T $=$ 10 & T $=$ N \\\\\n\\midrule\n1 ? N $<$ 10 & 21.8 & 22.4 & 22.4 \\\\\n10 ? N $<$ 20 & 24.6 & 25.9 & 26 \\\\\n20 ? N $<$ 30 & 24.9 & 26.7 & 27.1 \\\\\n30 ? N $<$ 40 & 24.9 & 26.7 & 27.6 \\\\\n40 ? N & 25 & 27.5 & 28.1 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: The performance (BLEU) of base CMLM with different amounts of mask-predict iterations (T ) on WMT\u201914 EN-DE, bucketed by target sequence length (N). Decoding with (cid:96) = 1 length candidates.\n
\n \n T = 4\n \n T = 10\n \n T = N\n
\n 1 ? N < 10\n \n 21.8\n \n 22.4\n \n 22.4\n
\n 10 ? N < 20\n \n 24.6\n \n 25.9\n \n 26\n
\n 20 ? N < 30\n \n 24.9\n \n 26.7\n \n 27.1\n
\n 30 ? N < 40\n \n 24.9\n \n 26.7\n \n 27.6\n
\n 40 ? N\n \n 25\n \n 27.5\n \n 28.1\n
\n" }, { "instance_id": "57cca0ef-5fdd-442e-a6e6-297015cf6f3f", "table_id_paper": "D19-1674table_2", "caption": "Human evaluation results. The values in parentheses are standard deviations.", "row_header_level": 2, "row_headers": [ [ "Method", "Proposed" ], [ "Method", "Baseline" ] ], "column_header_level": 1, "column_headers": [ [ "Average" ], [ "Score" ] ], "contents": [ [ "4.11", "(0.99)" ], [ "2.18", "(1.17)" ] ], "metrics_loc": "column", "metrics_type": "['Average', 'Score']", "target_entity": "['Proposed']", "table_html": "
Average Score
Method || Proposed 4.11 (0.99)
Method || Baseline 2.18 (1.17)
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "D19-1674", "page_no": 4, "dir": "emnlp2019", "description": "Table 2 shows the human evaluation results, where we compute the average score by first computing the average score of the three evaluators for each of the top-K outputs, then selecting the maximum of the K scores as the score of the method. The average shown in the table is the average of such maximum scores. The proposed method failed to generate any anagram for 3 instances. We therefore set the score of the proposed method to 1.0 for them. We can see that the proposed method yielded significantly higher human evaluation scores. Table 3 shows example anagrams generated by the proposed method and the baseline method. While the baseline method seems to employ some heuristics for deciding the word order, the combinations of selected words tend to form meaningless sentences. On the other hand, the proposed method tends to generate readable sentences. The anagrams shown in the table are listed in decreasing order of sentence probability. We can find that sentences with high probability are not always highly rated by the human evaluations. However, the normalized discounted cumulative gain (NDCG) score for the proposed method was 0.86. It suggests that using sentence probability works well for finding natural anagrams.", "class_sentence": "[1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0]", "sentences": "['Table 2 shows the human evaluation results, where we compute the average score by first computing the average score of the three evaluators for each of the top-K outputs, then selecting the maximum of the K scores as the score of the method. The average shown in the table is the average of such maximum scores.', 'The proposed method failed to generate any anagram for 3 instances.', 'We therefore set the score of the proposed method to 1.0 for them.', 'We can see that the proposed method yielded significantly higher human evaluation scores.', 'Table 3 shows example anagrams generated by the proposed method and the baseline method.', 'While the baseline method seems to employ some heuristics for deciding the word order, the combinations of selected words tend to form meaningless sentences.', 'On the other hand, the proposed method tends to generate readable sentences.', 'The anagrams shown in the table are listed in decreasing order of sentence probability.', 'We can find that sentences with high probability are not always highly rated by the human evaluations.', 'However, the normalized discounted cumulative gain (NDCG) score for the proposed method was 0.86.', 'It suggests that using sentence probability works well for finding natural anagrams.']", "header_mention": "[None, ['Proposed'], ['Score', 'Proposed'], ['Proposed', 'Average'], None, None, None, None, None, None, None]", "valid": 1, "image_id": "D19-1674-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Human evaluation results. The values in parentheses are standard deviations.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n AverageScore
Method || Proposed4.11(0.99)
Method || Baseline2.18(1.17)
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Human evaluation results. The values in parentheses are standard deviations.}\n\\begin{tabular}{lrl}\n\\toprule\n & Average & Score \\\\\n\\midrule\nMethod $\\parallel$ Proposed & 4.11 & (0.99) \\\\\nMethod $\\parallel$ Baseline & 2.18 & (1.17) \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Human evaluation results. The values in parentheses are standard deviations.\n
\n \n Average\n \n Score\n
\n Method || Proposed\n \n 4.11\n \n (0.99)\n
\n Method || Baseline\n \n 2.18\n \n (1.17)\n
\n" }, { "instance_id": "aa7eb32b-5aa2-46b0-b504-21dbc3147225", "table_id_paper": "P16-1165table_9", "caption": "Results of CRFs on CON dataset.", "row_header_level": 1, "row_headers": [ [ "MEb" ], [ "B-LSTMp" ], [ "MEe" ], [ "CRF (LC-NO)" ], [ "CRF (LC-LC)" ], [ "CRF (LC-LC1)" ], [ "CRF (LC-FC1)" ], [ "CRF (FC-FC)" ] ], "column_header_level": 1, "column_headers": [ [ "QC3" ], [ "TA" ] ], "contents": [ [ "56.67", "63.29" ], [ "65.15", "66.93" ], [ "59.94", "59.55" ], [ "62.20", "60.30" ], [ "62.35", "60.30" ], [ "65.94", "61.58" ], [ "61.18", "60.00" ], [ "64.54", "61.64" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy']", "target_entity": "['CRF (LC-NO)', 'CRF (LC-LC)', 'CRF (LC-LC1)', 'CRF (LC-FC1)', 'CRF (FC-FC)']", "table_html": "
QC3 TA
MEb 56.67 63.29
B-LSTMp 65.15 66.93
MEe 59.94 59.55
CRF (LC-NO) 62.20 60.30
CRF (LC-LC) 62.35 60.30
CRF (LC-LC1) 65.94 61.58
CRF (LC-FC1) 61.18 60.00
CRF (FC-FC) 64.54 61.64
", "table_name": "Table 9", "table_id": "table_9", "paper_id": "P16-1165", "page_no": 8, "dir": "acl2016", "description": "As baselines, we use three models: (i) MEb, (ii) Ba MaxEnt using BOW representation; LSTMp, which is now trained on the concatenated set of sentences from MRDA and CON training sets; and (iii) MEe, a MaxEnt using sentence embeddings extracted from the B-LSTMp, i.e., the sentence embeddings are used as feature vectors. We experiment with the CRF variants in Table 1. The CRFs are trained on the CON training set using the sentence embeddings that are extracted by applying the B-LSTMp model, as was done with MEe. Table 9 shows our results. We notice that CRFs generally outperform MEs in accuracy. This indicates that there are conversational dependencies between the sentences in a conversation. When we compare between CRF variants, we notice that the model that does not consider any link across comments perform the worst; see CRF (LC-NO). A simple linear chain connection between sentences in their temporal order does not improve much (CRF (LC-LC)), which indicates that the widely used linear chain CRF (Lafferty et al., 2001) is not the most appropriate model for capturing conversational dependencies in these conversations. The CRF (LC-LC1) is one of the best performing models and perform significantly better than B-LSTMp. This model considers linear chain connections between sentences inside comments and only to the first comment. Note that both QC3 and TA are forum sites, where participants in a conversation interact mostly with the person who posts the first comment asking for some information. This is interesting that our model can capture this aspect. Another interesting observation is that when we change the above model to consider relations with every sentence in the first comment (CRF (LC-FC1)), this degrades the performance. This could be due to the fact that the information seeking person first explains her situation, and then asks for the information. Others tend to respond to the requested information rather than to her situation. The CRF (FC-FC) also yields as good results as CRF (LC-LC1).", "class_sentence": "[2, 2, 2, 1, 2, 2, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2]", "sentences": "['As baselines, we use three models: (i) MEb, (ii) Ba MaxEnt using BOW representation; LSTMp, which is now trained on the concatenated set of sentences from MRDA and CON training sets; and (iii) MEe, a MaxEnt using sentence embeddings extracted from the B-LSTMp, i.e., the sentence embeddings are used as feature vectors.', 'We experiment with the CRF variants in Table 1.', 'The CRFs are trained on the CON training set using the sentence embeddings that are extracted by applying the B-LSTMp model, as was done with MEe.', 'Table 9 shows our results.', 'We notice that CRFs generally outperform MEs in accuracy.', 'This indicates that there are conversational dependencies between the sentences in a conversation.', 'When we compare between CRF variants, we notice that the model that does not consider any link across comments perform the worst; see CRF (LC-NO).', 'A simple linear chain connection between sentences in their temporal order does not improve much (CRF (LC-LC)), which indicates that the widely used linear chain CRF (Lafferty et al., 2001) is not the most appropriate model for capturing conversational dependencies in these conversations.', 'The CRF (LC-LC1) is one of the best performing models and perform significantly better than B-LSTMp.', 'This model considers linear chain connections between sentences inside comments and only to the first comment.', 'Note that both QC3 and TA are forum sites, where participants in a conversation interact mostly with the person who posts the first comment asking for some information.', 'This is interesting that our model can capture this aspect.', 'Another interesting observation is that when we change the above model to consider relations with every sentence in the first comment (CRF (LC-FC1)), this degrades the performance.', 'This could be due to the fact that the information seeking person first explains her situation, and then asks for the information.', 'Others tend to respond to the requested information rather than to her situation.', 'The CRF (FC-FC) also yields as good results as CRF (LC-LC1).']", "header_mention": "[['MEb', 'B-LSTMp', 'MEe'], ['CRF (LC-NO)', 'CRF (LC-LC)', 'CRF (LC-LC1)', 'CRF (LC-FC1)', 'CRF (FC-FC)'], ['CRF (LC-NO)', 'CRF (LC-LC)', 'CRF (LC-LC1)', 'CRF (LC-FC1)', 'CRF (FC-FC)'], None, ['CRF (LC-NO)', 'CRF (LC-LC)', 'CRF (LC-LC1)', 'CRF (LC-FC1)', 'CRF (FC-FC)', 'MEb', 'MEe'], None, ['CRF (LC-NO)'], ['CRF (LC-LC)'], ['CRF (LC-LC1)', 'B-LSTMp'], ['CRF (LC-LC1)'], ['QC3', 'TA'], ['QC3', 'TA'], None, None, None, None]", "valid": 1, "image_id": "P16-1165-Table9-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results of CRFs on CON dataset.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n QC3TA
MEb56.6763.29
B-LSTMp65.1566.93
MEe59.9459.55
CRF (LC-NO)62.2060.30
CRF (LC-LC)62.3560.30
CRF (LC-LC1)65.9461.58
CRF (LC-FC1)61.1860.00
CRF (FC-FC)64.5461.64
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{8}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results of CRFs on CON dataset.}\n\\begin{tabular}{lrr}\n\\toprule\n & QC3 & TA \\\\\n\\midrule\nMEb & 56.67 & 63.29 \\\\\nB-LSTMp & 65.15 & 66.93 \\\\\nMEe & 59.94 & 59.55 \\\\\nCRF (LC-NO) & 62.2 & 60.3 \\\\\nCRF (LC-LC) & 62.35 & 60.3 \\\\\nCRF (LC-LC1) & 65.94 & 61.58 \\\\\nCRF (LC-FC1) & 61.18 & 60 \\\\\nCRF (FC-FC) & 64.54 & 61.64 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 9: Results of CRFs on CON dataset.\n
\n \n QC3\n \n TA\n
\n MEb\n \n 56.67\n \n 63.29\n
\n B-LSTMp\n \n 65.15\n \n 66.93\n
\n MEe\n \n 59.94\n \n 59.55\n
\n CRF (LC-NO)\n \n 62.20\n \n 60.30\n
\n CRF (LC-LC)\n \n 62.35\n \n 60.30\n
\n CRF (LC-LC1)\n \n 65.94\n \n 61.58\n
\n CRF (LC-FC1)\n \n 61.18\n \n 60.00\n
\n CRF (FC-FC)\n \n 64.54\n \n 61.64\n
\n" }, { "instance_id": "8cc6debe-8422-4bae-8bd4-a31544ecedde", "table_id_paper": "P16-1181table_4", "caption": "Parsing results (LAS) on test sets for different sentence boundaries.", "row_header_level": 1, "row_headers": [ [ "GOLD" ], [ "MARMOT" ], [ "NOSYNTAX" ], [ "JOINT" ], [ "JOINT-REPARSED" ] ], "column_header_level": 1, "column_headers": [ [ "WSJ" ], [ "Switchboard" ], [ "WSJ*" ] ], "contents": [ [ "90.22", "84.99", "88.71" ], [ "89.81", "78.93", "83.37" ], [ "89.95", "80.30\u2020", "83.61" ], [ "89.71", "79.97\u2020", "85.66\u2020\u2021" ], [ "89.93", "80.61\u2020\u2021\u2217", "85.38\u2020\u2021" ] ], "metrics_loc": "column", "metrics_type": "['LAS', 'LAS', 'LAS']", "target_entity": "['GOLD', 'JOINT', 'JOINT-REPARSED']", "table_html": "
WSJ Switchboard WSJ*
GOLD 90.22 84.99 88.71
MARMOT 89.81 78.93 83.37
NOSYNTAX 89.95 80.30\u2020 83.61
JOINT 89.71 79.97\u2020 85.66\u2020\u2021
JOINT-REPARSED 89.93 80.61\u2020\u2021\u2217 85.38\u2020\u2021
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P16-1181", "page_no": 8, "dir": "acl2016", "description": "Table 4 shows the results on the test sets. For WSJ, where sentence segmentation is almost trivial, we see only minor drops in LAS between GOLD and the systems that use predicted sentence boundaries. Among the systems that use predicted boundaries, no differences are significant. For WSJ\u2217 and Switchboard the picture is much different. Compared to GOLD, all systems show considerable drops in accuracy which asserts that errors from the sentence boundary detection task propagate to the parser and worsen the parser accuracy. On Switchboard the parsers yield significantly better results than MARMOT. The best result is obtained after reparsing and this is also significantly better than any other system. Although there is a slight drop in accuracy between NOSYNTAX and JOINT, this difference is not significant. The results on WSJ\u2217 show that not only does syntax help to improve sentence segmentation, it does so to a degree that parsing results deteriorate when simpler sentence boundary detectors are used. Here, both JOINT and JOINT-REPARSED obtain significantly better parsing accuracies than the systems that do not have access to syntax during sentence boundary prediction. Although JOINT-REPARSED performs a bit worse, the difference compared to JOINT is not significant.", "class_sentence": "[1, 1, 2, 1, 1, 1, 1, 1, 2, 1, 1]", "sentences": "['Table 4 shows the results on the test sets.', 'For WSJ, where sentence segmentation is almost trivial, we see only minor drops in LAS between GOLD and the systems that use predicted sentence boundaries.', 'Among the systems that use predicted boundaries, no differences are significant.', 'For WSJ\u2217 and Switchboard the picture is much different.', 'Compared to GOLD, all systems show considerable drops in accuracy which asserts that errors from the sentence boundary detection task propagate to the parser and worsen the parser accuracy.', 'On Switchboard the parsers yield significantly better results than MARMOT.', 'The best result is obtained after reparsing and this is also significantly better than any other system.', 'Although there is a slight drop in accuracy between NOSYNTAX and JOINT, this difference is not significant.', 'The results on WSJ\u2217 show that not only does syntax help to improve sentence segmentation, it does so to a degree that parsing results deteriorate when simpler sentence boundary detectors are used.', 'Here, both JOINT and JOINT-REPARSED obtain significantly better parsing accuracies than the systems that do not have access to syntax during sentence boundary prediction.', 'Although JOINT-REPARSED performs a bit worse, the difference compared to JOINT is not significant.']", "header_mention": "[None, ['WSJ', 'GOLD'], None, ['WSJ*', 'Switchboard'], ['GOLD'], ['Switchboard', 'GOLD', 'MARMOT'], ['Switchboard', 'JOINT-REPARSED'], ['Switchboard', 'JOINT-REPARSED', 'NOSYNTAX', 'JOINT'], ['WSJ*'], ['WSJ*', 'JOINT-REPARSED', 'JOINT'], ['WSJ*', 'JOINT-REPARSED', 'JOINT']]", "valid": 1, "image_id": "P16-1181-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Parsing results (LAS) on test sets for different sentence boundaries.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n WSJSwitchboardWSJ*
GOLD90.2284.9988.71
MARMOT89.8178.9383.37
NOSYNTAX89.9580.30\u202083.61
JOINT89.7179.97\u202085.66\u2020\u2021
JOINT-REPARSED89.9380.61\u2020\u2021\u221785.38\u2020\u2021
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Parsing results (LAS) on test sets for different sentence boundaries.}\n\\begin{tabular}{lrll}\n\\toprule\n & WSJ & Switchboard & WSJ* \\\\\n\\midrule\nGOLD & 90.22 & 84.99 & 88.71 \\\\\nMARMOT & 89.81 & 78.93 & 83.37 \\\\\nNOSYNTAX & 89.95 & 80.30$\\dagger$ & 83.61 \\\\\nJOINT & 89.71 & 79.97$\\dagger$ & 85.66$\\dagger$$\\ddagger$ \\\\\nJOINT-REPARSED & 89.93 & 80.61$\\dagger$$\\ddagger$$\\ast$ & 85.38$\\dagger$$\\ddagger$ \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Parsing results (LAS) on test sets for different sentence boundaries.\n
\n \n WSJ\n \n Switchboard\n \n WSJ*\n
\n GOLD\n \n 90.22\n \n 84.99\n \n 88.71\n
\n MARMOT\n \n 89.81\n \n 78.93\n \n 83.37\n
\n NOSYNTAX\n \n 89.95\n \n 80.30\u2020\n \n 83.61\n
\n JOINT\n \n 89.71\n \n 79.97\u2020\n \n 85.66\u2020\u2021\n
\n JOINT-REPARSED\n \n 89.93\n \n 80.61\u2020\u2021\u2217\n \n 85.38\u2020\u2021\n
\n" }, { "instance_id": "27d04574-e260-4f18-b786-eb9f4327011c", "table_id_paper": "P16-1189table_4", "caption": "Best F1 measures on the Accurat evaluation sets", "row_header_level": 4, "row_headers": [ [ "SYSTEM", "LEXACC", "TEST SETS", "1:1" ], [ "SYSTEM", "STACC", "TEST SETS", "1:1" ], [ "SYSTEM", "LEXACC", "TEST SETS", "2:1" ], [ "SYSTEM", "STACC", "TEST SETS", "2:1" ], [ "SYSTEM", "LEXACC", "TEST SETS", "100:1" ], [ "SYSTEM", "STACC", "TEST SETS", "100:1" ] ], "column_header_level": 1, "column_headers": [ [ "EN-DE" ], [ "EN-EL" ], [ "EN-ET" ], [ "EN-LT" ], [ "EN-LV" ], [ "EN-RO" ], [ "EN-SL" ] ], "contents": [ [ "96.0", "89.5", "88.9", "93.1", "95.0", "99.4", "88.5" ], [ "96.7", "88.0", "92.0", "96.1", "96.6", "98.8", "89.5" ], [ "83.4", "83.2", "73.9", "81.2", "83.8", "95.3", "81.6" ], [ "89.2", "83.2", "79.9", "86.9", "88.2", "95.3", "82.3" ], [ "16.6", "22.7", "34.2", "45.1", "45.1", "70.4", "24.9" ], [ "33.7", "37.3", "42.5", "56.0", "56.2", "75.7", "35.3" ] ], "metrics_loc": "column", "metrics_type": "['F1', 'F1', 'F1', 'F1', 'F1', 'F1', 'F1']", "target_entity": "['LEXACC', 'STACC']", "table_html": "
EN-DE EN-EL EN-ET EN-LT EN-LV EN-RO EN-SL
SYSTEM || LEXACC || TEST SETS || 1:1 96.0 89.5 88.9 93.1 95.0 99.4 88.5
SYSTEM || STACC || TEST SETS || 1:1 96.7 88.0 92.0 96.1 96.6 98.8 89.5
SYSTEM || LEXACC || TEST SETS || 2:1 83.4 83.2 73.9 81.2 83.8 95.3 81.6
SYSTEM || STACC || TEST SETS || 2:1 89.2 83.2 79.9 86.9 88.2 95.3 82.3
SYSTEM || LEXACC || TEST SETS || 100:1 16.6 22.7 34.2 45.1 45.1 70.4 24.9
SYSTEM || STACC || TEST SETS || 100:1 33.7 37.3 42.5 56.0 56.2 75.7 35.3
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P16-1189", "page_no": 7, "dir": "acl2016", "description": "Table 4 presents the results on the Accurat test sets for LEXACC and STACC using their respective optimal similarity thresholds. On the 21 test sets, the two systems were tied on two occasions, with STACC obtaining better results in 89.5% of the remaining cases. On the noisiest datasets, STACC was consistently and markedly better across language pairs.", "class_sentence": "[1, 1, 1]", "sentences": "['Table 4 presents the results on the Accurat test sets for LEXACC and STACC using their respective optimal similarity thresholds.', 'On the 21 test sets, the two systems were tied on two occasions, with STACC obtaining better results in 89.5% of the remaining cases.', 'On the noisiest datasets, STACC was consistently and markedly better across language pairs.']", "header_mention": "[['LEXACC', 'STACC'], ['STACC'], ['STACC']]", "valid": 1, "image_id": "P16-1189-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Best F1 measures on the Accurat evaluation sets

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n EN-DEEN-ELEN-ETEN-LTEN-LVEN-ROEN-SL
SYSTEM || LEXACC || TEST SETS || 1:196.089.588.993.195.099.488.5
SYSTEM || STACC || TEST SETS || 1:196.788.092.096.196.698.889.5
SYSTEM || LEXACC || TEST SETS || 2:183.483.273.981.283.895.381.6
SYSTEM || STACC || TEST SETS || 2:189.283.279.986.988.295.382.3
SYSTEM || LEXACC || TEST SETS || 100:116.622.734.245.145.170.424.9
SYSTEM || STACC || TEST SETS || 100:133.737.342.556.056.275.735.3
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Best F1 measures on the Accurat evaluation sets}\n\\begin{tabular}{lrrrrrrr}\n\\toprule\n & EN-DE & EN-EL & EN-ET & EN-LT & EN-LV & EN-RO & EN-SL \\\\\n\\midrule\nSYSTEM $\\parallel$ LEXACC $\\parallel$ TEST SETS $\\parallel$ 1:1 & 96 & 89.5 & 88.9 & 93.1 & 95 & 99.4 & 88.5 \\\\\nSYSTEM $\\parallel$ STACC $\\parallel$ TEST SETS $\\parallel$ 1:1 & 96.7 & 88 & 92 & 96.1 & 96.6 & 98.8 & 89.5 \\\\\nSYSTEM $\\parallel$ LEXACC $\\parallel$ TEST SETS $\\parallel$ 2:1 & 83.4 & 83.2 & 73.9 & 81.2 & 83.8 & 95.3 & 81.6 \\\\\nSYSTEM $\\parallel$ STACC $\\parallel$ TEST SETS $\\parallel$ 2:1 & 89.2 & 83.2 & 79.9 & 86.9 & 88.2 & 95.3 & 82.3 \\\\\nSYSTEM $\\parallel$ LEXACC $\\parallel$ TEST SETS $\\parallel$ 100:1 & 16.6 & 22.7 & 34.2 & 45.1 & 45.1 & 70.4 & 24.9 \\\\\nSYSTEM $\\parallel$ STACC $\\parallel$ TEST SETS $\\parallel$ 100:1 & 33.7 & 37.3 & 42.5 & 56 & 56.2 & 75.7 & 35.3 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Best F1 measures on the Accurat evaluation sets\n
\n \n EN-DE\n \n EN-EL\n \n EN-ET\n \n EN-LT\n \n EN-LV\n \n EN-RO\n \n EN-SL\n
\n SYSTEM || LEXACC || TEST SETS || 1:1\n \n 96.0\n \n 89.5\n \n 88.9\n \n 93.1\n \n 95.0\n \n 99.4\n \n 88.5\n
\n SYSTEM || STACC || TEST SETS || 1:1\n \n 96.7\n \n 88.0\n \n 92.0\n \n 96.1\n \n 96.6\n \n 98.8\n \n 89.5\n
\n SYSTEM || LEXACC || TEST SETS || 2:1\n \n 83.4\n \n 83.2\n \n 73.9\n \n 81.2\n \n 83.8\n \n 95.3\n \n 81.6\n
\n SYSTEM || STACC || TEST SETS || 2:1\n \n 89.2\n \n 83.2\n \n 79.9\n \n 86.9\n \n 88.2\n \n 95.3\n \n 82.3\n
\n SYSTEM || LEXACC || TEST SETS || 100:1\n \n 16.6\n \n 22.7\n \n 34.2\n \n 45.1\n \n 45.1\n \n 70.4\n \n 24.9\n
\n SYSTEM || STACC || TEST SETS || 100:1\n \n 33.7\n \n 37.3\n \n 42.5\n \n 56.0\n \n 56.2\n \n 75.7\n \n 35.3\n
\n" }, { "instance_id": "526cae64-4c04-4184-93d5-4899f1afa6e5", "table_id_paper": "P16-1201table_4", "caption": "Performance of the basic ED model. ANN uses pre-trained word embeddings while ANNRandom uses randomly initialized embeddings.", "row_header_level": 2, "row_headers": [ [ "Methods", "Nguyen\u2019s CNN (2015)" ], [ "Methods", "Chen\u2019s DMCNN (2015)" ], [ "Methods", "Liu\u2019s Approach (2016)" ], [ "Methods", "ANN (ours)" ], [ "Methods", "ANN-Random (ours)" ] ], "column_header_level": 1, "column_headers": [ [ "Pre" ], [ "Rec" ], [ "F1" ] ], "contents": [ [ "71.8", "66.4", "69.0" ], [ "75.6", "63.6", "69.1" ], [ "75.3", "64.4", "69.4" ], [ "79.5", "60.7", "68.8" ], [ "81.0", "49.5", "61.5" ] ], "metrics_loc": "column", "metrics_type": "['Pre', 'Rec', 'F1']", "target_entity": "['ANN (ours)', 'ANN-Random (ours)']", "table_html": "
Pre Rec F1
Methods || Nguyen\u2019s CNN (2015) 71.8 66.4 69.0
Methods || Chen\u2019s DMCNN (2015) 75.6 63.6 69.1
Methods || Liu\u2019s Approach (2016) 75.3 64.4 69.4
Methods || ANN (ours) 79.5 60.7 68.8
Methods || ANN-Random (ours) 81.0 49.5 61.5
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P16-1201", "page_no": 6, "dir": "acl2016", "description": "Table 4 shows the experimental results, from which we can see that the three-layer ANN model is surprisingly effective for event detection, which even yields competitive results compared with Nguyen\u2019s CNN and Chen\u2019s DMCNN. We believe the reason is that, compared with CNN and DMCNN, ANN focuses on capturing lexical features which have been proved much more important than sentence features for the ED task by (Chen et al., 2015).,. Moreover, our basic model achieves much higher precision than state-of-the-art approaches (79.5% vs. 75.6%). We also investigate the performance of the basic ED model without pre-trained word embeddings (denoted by ANN-Random). The result shows that randomly initialized word embeddings decrease the F1 score by 7.3 (61.5 vs. 68.8).", "class_sentence": "[1, 2, 1, 2, 1]", "sentences": "['Table 4 shows the experimental results, from which we can see that the three-layer ANN model is surprisingly effective for event detection, which even yields competitive results compared with Nguyen\u2019s CNN and Chen\u2019s DMCNN.', 'We believe the reason is that, compared with CNN and DMCNN, ANN focuses on capturing lexical features which have been proved much more important than sentence features for the ED task by (Chen et al., 2015).,.', 'Moreover, our basic model achieves much higher precision than state-of-the-art approaches (79.5% vs. 75.6%).', 'We also investigate the performance of the basic ED model without pre-trained word embeddings (denoted by ANN-Random).', 'The result shows that randomly initialized word embeddings decrease the F1 score by 7.3 (61.5 vs. 68.8).']", "header_mention": "[['ANN (ours)', 'Nguyen\u2019s CNN (2015)', 'Chen\u2019s DMCNN (2015)'], ['ANN (ours)', 'Nguyen\u2019s CNN (2015)', 'Chen\u2019s DMCNN (2015)'], ['ANN (ours)', 'Nguyen\u2019s CNN (2015)'], ['ANN-Random (ours)'], ['ANN-Random (ours)', 'ANN (ours)']]", "valid": 1, "image_id": "P16-1201-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance of the basic ED model. ANN uses pre-trained word embeddings while ANNRandom uses randomly initialized embeddings.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n PreRecF1
Methods || Nguyen\u2019s CNN (2015)71.866.469.0
Methods || Chen\u2019s DMCNN (2015)75.663.669.1
Methods || Liu\u2019s Approach (2016)75.364.469.4
Methods || ANN (ours)79.560.768.8
Methods || ANN-Random (ours)81.049.561.5
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance of the basic ED model. ANN uses pre-trained word embeddings while ANNRandom uses randomly initialized embeddings.}\n\\begin{tabular}{lrrr}\n\\toprule\n & Pre & Rec & F1 \\\\\n\\midrule\nMethods $\\parallel$ Nguyen\u2019s CNN (2015) & 71.8 & 66.4 & 69 \\\\\nMethods $\\parallel$ Chen\u2019s DMCNN (2015) & 75.6 & 63.6 & 69.1 \\\\\nMethods $\\parallel$ Liu\u2019s Approach (2016) & 75.3 & 64.4 & 69.4 \\\\\nMethods $\\parallel$ ANN (ours) & 79.5 & 60.7 & 68.8 \\\\\nMethods $\\parallel$ ANN-Random (ours) & 81 & 49.5 & 61.5 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Performance of the basic ED model. ANN uses pre-trained word embeddings while ANNRandom uses randomly initialized embeddings.\n
\n \n Pre\n \n Rec\n \n F1\n
\n Methods || Nguyen\u2019s CNN (2015)\n \n 71.8\n \n 66.4\n \n 69.0\n
\n Methods || Chen\u2019s DMCNN (2015)\n \n 75.6\n \n 63.6\n \n 69.1\n
\n Methods || Liu\u2019s Approach (2016)\n \n 75.3\n \n 64.4\n \n 69.4\n
\n Methods || ANN (ours)\n \n 79.5\n \n 60.7\n \n 68.8\n
\n Methods || ANN-Random (ours)\n \n 81.0\n \n 49.5\n \n 61.5\n
\n" }, { "instance_id": "4970de49-7086-4d4d-a26c-18e4c5f1c373", "table_id_paper": "P16-1201table_5", "caption": "Results of manual evaluations.", "row_header_level": 3, "row_headers": [ [ "Baselines", "Methods", "ANN" ], [ "Baselines", "Methods", "SF" ], [ "Baselines", "Methods", "RF" ], [ "Baselines", "Methods", "SL" ], [ "-", "Methods", "PSL-based Approach" ] ], "column_header_level": 1, "column_headers": [ [ "Precision (%)" ] ], "contents": [ [ "77.5" ], [ "72.0" ], [ "71.0" ], [ "79.5" ], [ "81.0" ] ], "metrics_loc": "column", "metrics_type": "['Precision (%)']", "target_entity": "['PSL-based Approach']", "table_html": "
Precision (%)
Baselines || Methods || ANN 77.5
Baselines || Methods || SF 72.0
Baselines || Methods || RF 71.0
Baselines || Methods || SL 79.5
- || Methods || PSL-based Approach 81.0
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "P16-1201", "page_no": 7, "dir": "acl2016", "description": "Table 5 shows the results of manual evaluations. Through the comparison of ANN and SF, we can see that the application of H1 caused a loss of 5.5 point. It happens mainly because the performance of SF is very sensitive to the wrongly mapped frames. That is, if a frame is mismapped, then all sentences under it would be mislabeled as events. Thus, even a single mismapped frame could significantly hurt the performance. This result also proves that H1 is inappropriate to be used as a hard constraint. As H2 is only an extension of H1, RF performs similarly with SF. Moreover, SL obtains a gain of 2.0% improvement compared with ANN, which demonstrates that the \u201dsame LU\u201d hypothesis is very useful. Finally, with all the hypotheses, the PSL-based approach achieves the best performance, which demonstrates that our hypotheses are useful and it is an effective way to jointly utilize them as soft constraints through PSL for event detection in FN.", "class_sentence": "[1, 1, 2, 2, 2, 2, 2, 1, 1]", "sentences": "['Table 5 shows the results of manual evaluations.', 'Through the comparison of ANN and SF, we can see that the application of H1 caused a loss of 5.5 point.', 'It happens mainly because the performance of SF is very sensitive to the wrongly mapped frames.', 'That is, if a frame is mismapped, then all sentences under it would be mislabeled as events.', 'Thus, even a single mismapped frame could significantly hurt the performance.', 'This result also proves that H1 is inappropriate to be used as a hard constraint.', 'As H2 is only an extension of H1, RF performs similarly with SF.', 'Moreover, SL obtains a gain of 2.0% improvement compared with ANN, which demonstrates that the \u201dsame LU\u201d hypothesis is very useful.', 'Finally, with all the hypotheses, the PSL-based approach achieves the best performance, which demonstrates that our hypotheses are useful and it is an effective way to jointly utilize them as soft constraints through PSL for event detection in FN.']", "header_mention": "[None, ['ANN', 'SF'], ['SF'], None, None, None, None, ['SL', 'ANN'], ['PSL-based Approach']]", "valid": 1, "image_id": "P16-1201-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results of manual evaluations.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Precision (%)
Baselines || Methods || ANN77.5
Baselines || Methods || SF72.0
Baselines || Methods || RF71.0
Baselines || Methods || SL79.5
- || Methods || PSL-based Approach81.0
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results of manual evaluations.}\n\\begin{tabular}{lr}\n\\toprule\n & Precision (\\%) \\\\\n\\midrule\nBaselines $\\parallel$ Methods $\\parallel$ ANN & 77.5 \\\\\nBaselines $\\parallel$ Methods $\\parallel$ SF & 72 \\\\\nBaselines $\\parallel$ Methods $\\parallel$ RF & 71 \\\\\nBaselines $\\parallel$ Methods $\\parallel$ SL & 79.5 \\\\\n- $\\parallel$ Methods $\\parallel$ PSL-based Approach & 81 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Results of manual evaluations.\n
\n \n Precision (%)\n
\n Baselines || Methods || ANN\n \n 77.5\n
\n Baselines || Methods || SF\n \n 72.0\n
\n Baselines || Methods || RF\n \n 71.0\n
\n Baselines || Methods || SL\n \n 79.5\n
\n - || Methods || PSL-based Approach\n \n 81.0\n
\n" }, { "instance_id": "f504740b-2e9c-4884-ae68-d25126a17a48", "table_id_paper": "P16-2011table_2", "caption": "Comparison of different methods on English event detection.", "row_header_level": 2, "row_headers": [ [ "Model", "MaxEnt" ], [ "Model", "Cross-Event" ], [ "Model", "Cross-Entity" ], [ "Model", "Joint Model" ], [ "Model", "PR" ], [ "Model", "CNN" ], [ "Model", "RNN" ], [ "Model", "LSTM" ], [ "Model", "Bi-LSTM" ], [ "Model", "HNN" ] ], "column_header_level": 2, "column_headers": [ [ "Trigger Identification", "P" ], [ "Trigger Identification", "R" ], [ "Trigger Identification", "F" ], [ "Trigger Classification", "P" ], [ "Trigger Classification", "R" ], [ "Trigger Classification", "F" ] ], "contents": [ [ "76.2", "60.5", "67.4", "74.5", "59.1", "65.9" ], [ "N/A", "N/A", "N/A", "68.7", "68.9", "68.8" ], [ "N/A", "N/A", "N/A", "72.9", "64.3", "68.3" ], [ "76.9", "65.0", "70.4", "73.7", "62.3", "67.5" ], [ "N/A", "N/A", "N/A", "68.9", "72.0", "70.4" ], [ "80.4", "67.7", "73.5", "75.6", "63.6", "69.1" ], [ "73.2", "63.5", "67.4", "67.3", "59.9", "64.2" ], [ "78.6", "67.4", "72.6", "74.5", "60.7", "66.9" ], [ "80.1", "69.4", "74.3", "81.6", "62.3", "70.6" ], [ "80.8", "71.5", "75.9", "84.6", "64.9", "73.4" ] ], "metrics_loc": "column", "metrics_type": "['P', 'R', 'F', 'P', 'R', 'F']", "target_entity": "['HNN']", "table_html": "
Trigger Identification || P Trigger Identification || R Trigger Identification || F Trigger Classification || P Trigger Classification || R Trigger Classification || F
Model || MaxEnt 76.2 60.5 67.4 74.5 59.1 65.9
Model || Cross-Event N/A N/A N/A 68.7 68.9 68.8
Model || Cross-Entity N/A N/A N/A 72.9 64.3 68.3
Model || Joint Model 76.9 65.0 70.4 73.7 62.3 67.5
Model || PR N/A N/A N/A 68.9 72.0 70.4
Model || CNN 80.4 67.7 73.5 75.6 63.6 69.1
Model || RNN 73.2 63.5 67.4 67.3 59.9 64.2
Model || LSTM 78.6 67.4 72.6 74.5 60.7 66.9
Model || Bi-LSTM 80.1 69.4 74.3 81.6 62.3 70.6
Model || HNN 80.8 71.5 75.9 84.6 64.9 73.4
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P16-2011", "page_no": 4, "dir": "acl2016", "description": "Table 2 shows the overall performance of all methods on the ACE2005 English corpus. We can see that our approach significantly outperforms all previous methods. The better performance of HNN can be further explained by the following reasons: (1) Compared with feature based methods, such as MaxEnt, Cross-Event, Cross-Entity, and Joint Model, neural network based methods (including CNN, Bi-LSTM, HNN) performs better because they can make better use of word semantic information and avoid the errors propagated from NLP tools which may hinder the performance for event detection. (2) Moreover, Bi-LSTM can capture both preceding and following sequence information, which is much richer than dependency path. For example, in S2, the semantic of \u201ccourt\u201d can be delivered to release by a forward sequence in our approach. It is an important clue which can help to predict \u201crelease\u201d as a trigger for \u201cReleaseParole\u201d. For explicit feature based methods, they can not establish a relation between \u201ccourt\u201d and \u201crelease\u201d, because they belong to different clauses, and there is no direct dependency path between them. While in our approach, the semantics of \u201ccourt\u201d can be delivered to release by a forward sequence. (3) Cross-entity system achieves higher recall because it uses not only sentence-level information but also document-level information. It utilizes event concordance to predict a local trigger\u2019s event type based on cross-sentence inference. For example, an \u201cattack\u201d event is more likely to occur with \u201ckilled\u201d or \u201cdie\u201d event rather than \u201cmarry\u201d event. However, this method heavily relies on lexical and syntactic features, thus the precision is lower than neural network based methods. (4) RNN and LSTM perform slightly worse than Bi-LSTM. An obvious reason is that RNN and LSTM only consider the preceding sequence information of the trigger, which may miss some important following clues. Considering S1 again, when extracting the trigger \u201creleases\u201d, both models will miss the following sequence \u201c20 million euros to Iraq\u201d. This may seriously hinder the performance of RNN and LSTM for event detection.", "class_sentence": "[1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2]", "sentences": "['Table 2 shows the overall performance of all methods on the ACE2005 English corpus.', 'We can see that our approach significantly outperforms all previous methods.', 'The better performance of HNN can be further explained by the following reasons: (1) Compared with feature based methods, such as MaxEnt, Cross-Event, Cross-Entity, and Joint Model, neural network based methods (including CNN, Bi-LSTM, HNN) performs better because they can make better use of word semantic information and avoid the errors propagated from NLP tools which may hinder the performance for event detection.', '(2) Moreover, Bi-LSTM can capture both preceding and following sequence information, which is much richer than dependency path.', 'For example, in S2, the semantic of \u201ccourt\u201d can be delivered to release by a forward sequence in our approach.', 'It is an important clue which can help to predict \u201crelease\u201d as a trigger for \u201cReleaseParole\u201d.', 'For explicit feature based methods, they can not establish a relation between \u201ccourt\u201d and \u201crelease\u201d, because they belong to different clauses, and there is no direct dependency path between them.', 'While in our approach, the semantics of \u201ccourt\u201d can be delivered to release by a forward sequence.', '(3) Cross-entity system achieves higher recall because it uses not only sentence-level information but also document-level information.', 'It utilizes event concordance to predict a local trigger\u2019s event type based on cross-sentence inference.', 'For example, an \u201cattack\u201d event is more likely to occur with \u201ckilled\u201d or \u201cdie\u201d event rather than \u201cmarry\u201d event.', 'However, this method heavily relies on lexical and syntactic features, thus the precision is lower than neural network based methods.', '(4) RNN and LSTM perform slightly worse than Bi-LSTM.', 'An obvious reason is that RNN and LSTM only consider the preceding sequence information of the trigger, which may miss some important following clues.', 'Considering S1 again, when extracting the trigger \u201creleases\u201d, both models will miss the following sequence \u201c20 million euros to Iraq\u201d.', 'This may seriously hinder the performance of RNN and LSTM for event detection.']", "header_mention": "[None, ['HNN'], ['HNN', 'MaxEnt', 'Cross-Event', 'Cross-Entity', 'Joint Model', 'CNN', 'Bi-LSTM'], None, None, None, None, None, ['Cross-Entity'], ['Cross-Entity'], ['Cross-Entity'], ['Cross-Entity'], ['RNN', 'LSTM', 'Bi-LSTM'], ['RNN', 'LSTM'], None, ['RNN', 'LSTM']]", "valid": 1, "image_id": "P16-2011-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison of different methods on English event detection.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Trigger Identification || PTrigger Identification || RTrigger Identification || FTrigger Classification || PTrigger Classification || RTrigger Classification || F
Model || MaxEnt76.260.567.474.559.165.9
Model || Cross-EventN/AN/AN/A68.768.968.8
Model || Cross-EntityN/AN/AN/A72.964.368.3
Model || Joint Model76.965.070.473.762.367.5
Model || PRN/AN/AN/A68.972.070.4
Model || CNN80.467.773.575.663.669.1
Model || RNN73.263.567.467.359.964.2
Model || LSTM78.667.472.674.560.766.9
Model || Bi-LSTM80.169.474.381.662.370.6
Model || HNN80.871.575.984.664.973.4
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison of different methods on English event detection.}\n\\begin{tabular}{llllrrr}\n\\toprule\n & Trigger Identification $\\parallel$ P & Trigger Identification $\\parallel$ R & Trigger Identification $\\parallel$ F & Trigger Classification $\\parallel$ P & Trigger Classification $\\parallel$ R & Trigger Classification $\\parallel$ F \\\\\n\\midrule\nModel $\\parallel$ MaxEnt & 76.2 & 60.5 & 67.4 & 74.5 & 59.1 & 65.9 \\\\\nModel $\\parallel$ Cross-Event & & & & 68.7 & 68.9 & 68.8 \\\\\nModel $\\parallel$ Cross-Entity & & & & 72.9 & 64.3 & 68.3 \\\\\nModel $\\parallel$ Joint Model & 76.9 & 65 & 70.4 & 73.7 & 62.3 & 67.5 \\\\\nModel $\\parallel$ PR & & & & 68.9 & 72 & 70.4 \\\\\nModel $\\parallel$ CNN & 80.4 & 67.7 & 73.5 & 75.6 & 63.6 & 69.1 \\\\\nModel $\\parallel$ RNN & 73.2 & 63.5 & 67.4 & 67.3 & 59.9 & 64.2 \\\\\nModel $\\parallel$ LSTM & 78.6 & 67.4 & 72.6 & 74.5 & 60.7 & 66.9 \\\\\nModel $\\parallel$ Bi-LSTM & 80.1 & 69.4 & 74.3 & 81.6 & 62.3 & 70.6 \\\\\nModel $\\parallel$ HNN & 80.8 & 71.5 & 75.9 & 84.6 & 64.9 & 73.4 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Comparison of different methods on English event detection.\n
\n \n Trigger Identification || P\n \n Trigger Identification || R\n \n Trigger Identification || F\n \n Trigger Classification || P\n \n Trigger Classification || R\n \n Trigger Classification || F\n
\n Model || MaxEnt\n \n 76.2\n \n 60.5\n \n 67.4\n \n 74.5\n \n 59.1\n \n 65.9\n
\n Model || Cross-Event\n \n N/A\n \n N/A\n \n N/A\n \n 68.7\n \n 68.9\n \n 68.8\n
\n Model || Cross-Entity\n \n N/A\n \n N/A\n \n N/A\n \n 72.9\n \n 64.3\n \n 68.3\n
\n Model || Joint Model\n \n 76.9\n \n 65.0\n \n 70.4\n \n 73.7\n \n 62.3\n \n 67.5\n
\n Model || PR\n \n N/A\n \n N/A\n \n N/A\n \n 68.9\n \n 72.0\n \n 70.4\n
\n Model || CNN\n \n 80.4\n \n 67.7\n \n 73.5\n \n 75.6\n \n 63.6\n \n 69.1\n
\n Model || RNN\n \n 73.2\n \n 63.5\n \n 67.4\n \n 67.3\n \n 59.9\n \n 64.2\n
\n Model || LSTM\n \n 78.6\n \n 67.4\n \n 72.6\n \n 74.5\n \n 60.7\n \n 66.9\n
\n Model || Bi-LSTM\n \n 80.1\n \n 69.4\n \n 74.3\n \n 81.6\n \n 62.3\n \n 70.6\n
\n Model || HNN\n \n 80.8\n \n 71.5\n \n 75.9\n \n 84.6\n \n 64.9\n \n 73.4\n
\n" }, { "instance_id": "829a4839-a663-4ff0-9321-badecf027866", "table_id_paper": "P16-2011table_4", "caption": "Results on Spanish event detection.", "row_header_level": 2, "row_headers": [ [ "Model", "LSTM" ], [ "Model", "Bi-LSTM" ], [ "Model", "HNN" ] ], "column_header_level": 2, "column_headers": [ [ "Trigger Identification", "P" ], [ "Trigger Identification", "R" ], [ "Trigger Identification", "F" ], [ "Trigger Classification", "P" ], [ "Trigger Classification", "R" ], [ "Trigger Classification", "F" ] ], "contents": [ [ "62.2", "52.9", "57.2", "56.9", "32.6", "41.6" ], [ "76.2", "63.1", "68.7", "61.5", "42.2", "50.1" ], [ "81.4", "65.2", "71.6", "66.3", "47.8", "55.5" ] ], "metrics_loc": "column", "metrics_type": "['P', 'R', 'F', 'P', 'R', 'F']", "target_entity": "['HNN']", "table_html": "
Trigger Identification || P Trigger Identification || R Trigger Identification || F Trigger Classification || P Trigger Classification || R Trigger Classification || F
Model || LSTM 62.2 52.9 57.2 56.9 32.6 41.6
Model || Bi-LSTM 76.2 63.1 68.7 61.5 42.2 50.1
Model || HNN 81.4 65.2 71.6 66.3 47.8 55.5
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P16-2011", "page_no": 5, "dir": "acl2016", "description": "Table 4 presents the performance of our method on the Spanish ERE corpus. The results show that HNN approach performed better than LSTM and Bi-LSTM. It indicates that our proposed model could achieve the best performance in multiple languages than other neural network methods. We did not compare our system with other systems (Tanev et al., 2009), because they reported the results on a non-standard data set .", "class_sentence": "[1, 1, 1, 2]", "sentences": "['Table 4 presents the performance of our method on the Spanish ERE corpus.', 'The results show that HNN approach performed better than LSTM and Bi-LSTM.', 'It indicates that our proposed model could achieve the best performance in multiple languages than other neural network methods.', 'We did not compare our system with other systems (Tanev et al., 2009), because they reported the results on a non-standard data set .']", "header_mention": "[None, ['HNN', 'LSTM', 'Bi-LSTM'], ['HNN'], None]", "valid": 1, "image_id": "P16-2011-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results on Spanish event detection.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Trigger Identification || PTrigger Identification || RTrigger Identification || FTrigger Classification || PTrigger Classification || RTrigger Classification || F
Model || LSTM62.252.957.256.932.641.6
Model || Bi-LSTM76.263.168.761.542.250.1
Model || HNN81.465.271.666.347.855.5
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results on Spanish event detection.}\n\\begin{tabular}{lrrrrrr}\n\\toprule\n & Trigger Identification $\\parallel$ P & Trigger Identification $\\parallel$ R & Trigger Identification $\\parallel$ F & Trigger Classification $\\parallel$ P & Trigger Classification $\\parallel$ R & Trigger Classification $\\parallel$ F \\\\\n\\midrule\nModel $\\parallel$ LSTM & 62.2 & 52.9 & 57.2 & 56.9 & 32.6 & 41.6 \\\\\nModel $\\parallel$ Bi-LSTM & 76.2 & 63.1 & 68.7 & 61.5 & 42.2 & 50.1 \\\\\nModel $\\parallel$ HNN & 81.4 & 65.2 & 71.6 & 66.3 & 47.8 & 55.5 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Results on Spanish event detection.\n
\n \n Trigger Identification || P\n \n Trigger Identification || R\n \n Trigger Identification || F\n \n Trigger Classification || P\n \n Trigger Classification || R\n \n Trigger Classification || F\n
\n Model || LSTM\n \n 62.2\n \n 52.9\n \n 57.2\n \n 56.9\n \n 32.6\n \n 41.6\n
\n Model || Bi-LSTM\n \n 76.2\n \n 63.1\n \n 68.7\n \n 61.5\n \n 42.2\n \n 50.1\n
\n Model || HNN\n \n 81.4\n \n 65.2\n \n 71.6\n \n 66.3\n \n 47.8\n \n 55.5\n
\n" }, { "instance_id": "ab67e7c9-7359-4f73-8ab4-d1a27d5c90b3", "table_id_paper": "P17-1034table_5", "caption": "SVM classification results across linguistic features (LF, bigrams here (Mukherjee et al., 2013b)), behavioral features (BF: RL, RD, MCS (Mukherjee et al., 2013b)); the SVM classification results by the intuitive method that finding the most similar existing review by edit distance ratio and take the found reviewers\u2019 behavioral features as approximation (BF EditSim+LF), and results by the intuitive method that finding the most similar existing review by averaged pre-trained word embeddings (using Word2Vec) (BF W2Vsim+W2V); and the SVM classification results across the learnt review embeddings (RE), the learnt review\u2019s rating embeddings (RRE), the learnt product\u2019s average rating embeddings (PRE) by our model. Improvements of our model are statistically significant with p<0.005 based on paired t-test.", "row_header_level": 2, "row_headers": [ [ "Features", "LF" ], [ "Features", "LF+BF" ], [ "Features", "BF EditSim+LF" ], [ "Features", "BF W2Vsim+W2V" ], [ "Features", "Ours RE" ], [ "Features", "Ours RE+RRE+PRE" ] ], "column_header_level": 2, "column_headers": [ [ "P", "Hotel" ], [ "P", "Restaurant" ], [ "R", "Hotel" ], [ "R", "Restaurant" ], [ "F1", "Hotel" ], [ "F1", "Restaurant" ], [ "A", "Hotel" ], [ "A", "Restaurant" ] ], "contents": [ [ "54.5", "53.8", "71.1", "80.8", "61.7", "64.6", "55.9", "55.8" ], [ "63.4", "58.1", "52.6", "61.2", "57.5", "59.6", "61.1", "58.5" ], [ "55.3", "53.9", "69.7", "82.82", "61.6", "65.1", "56.6", "56.0" ], [ "58.4", "56.3", "65.9", "73.4", "61.9", "63.7", "59.5", "58.2" ], [ "62.1", "58.4", "68.3", "75.1", "65.1", "65.7", "63.3", "60.8" ], [ "63.6", "59.0", "71.2", "78.8", "67.2", "67.5", "65.3", "62.0" ] ], "metrics_loc": "column", "metrics_type": "['P', 'P', 'R', 'R', 'F1', 'F1', 'A', 'A']", "target_entity": "['Ours RE', 'Ours RE+RRE+PRE']", "table_html": "
P || Hotel P || Restaurant R || Hotel R || Restaurant F1 || Hotel F1 || Restaurant A || Hotel A || Restaurant
Features || LF 54.5 53.8 71.1 80.8 61.7 64.6 55.9 55.8
Features || LF+BF 63.4 58.1 52.6 61.2 57.5 59.6 61.1 58.5
Features || BF EditSim+LF 55.3 53.9 69.7 82.82 61.6 65.1 56.6 56.0
Features || BF W2Vsim+W2V 58.4 56.3 65.9 73.4 61.9 63.7 59.5 58.2
Features || Ours RE 62.1 58.4 68.3 75.1 65.1 65.7 63.3 60.8
Features || Ours RE+RRE+PRE 63.6 59.0 71.2 78.8 67.2 67.5 65.3 62.0
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "P17-1034", "page_no": 7, "dir": "acl2017", "description": "To illustrate the effectiveness of our model, we conduct experiments on the public datasets, and make comparison with the most effective traditional linguistic features, e.g., bigrams, and the three practicable traditional behavioral features (RL, RD, MCS (Mukherjee et al., 2013b)) referred in Section 3.2. The results are shown in Table 5. For our model, we set the dimension of embeddings to 100, the number of CNN filters to 100, \u00ce\u00b8 to 0.1, Z to 2. The hyper-parameters are tuned by grid search on the development dataset. The product and reviewer embeddings are randomly intialized from a uniform distribution (Socher et al., 2013). The word embeddings are initialized with 100-dimensions vectors pre-trained by the CBOW model (Word2Vec) (Mikolov et al., 2013). As Table 5 showed, our model observably performs better in detecting review spam for the cold-start task in both hotel and restaurant domains.", "class_sentence": "[2, 1, 2, 2, 2, 2, 1]", "sentences": "['To illustrate the effectiveness of our model, we conduct experiments on the public datasets, and make comparison with the most effective traditional linguistic features, e.g., bigrams, and the three practicable traditional behavioral features (RL, RD, MCS (Mukherjee et al., 2013b)) referred in Section 3.2.', 'The results are shown in Table 5.', 'For our model, we set the dimension of embeddings to 100, the number of CNN filters to 100, \u00ce\u00b8 to 0.1, Z to 2.', 'The hyper-parameters are tuned by grid search on the development dataset.', 'The product and reviewer embeddings are randomly intialized from a uniform distribution (Socher et al., 2013).', 'The word embeddings are initialized with 100-dimensions vectors pre-trained by the CBOW model (Word2Vec) (Mikolov et al., 2013).', 'As Table 5 showed, our model observably performs better in detecting review spam for the cold-start task in both hotel and restaurant domains.']", "header_mention": "[['Ours RE', 'Ours RE+RRE+PRE'], None, ['Ours RE', 'Ours RE+RRE+PRE'], ['Ours RE', 'Ours RE+RRE+PRE'], None, ['Ours RE', 'Ours RE+RRE+PRE'], ['Ours RE', 'Ours RE+RRE+PRE']]", "valid": 1, "image_id": "P17-1034-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

SVM classification results across linguistic features (LF, bigrams here (Mukherjee et al., 2013b)), behavioral features (BF: RL, RD, MCS (Mukherjee et al., 2013b)); the SVM classification results by the intuitive method that finding the most similar existing review by edit distance ratio and take the found reviewers\u2019 behavioral features as approximation (BF EditSim+LF), and results by the intuitive method that finding the most similar existing review by averaged pre-trained word embeddings (using Word2Vec) (BF W2Vsim+W2V); and the SVM classification results across the learnt review embeddings (RE), the learnt review\u2019s rating embeddings (RRE), the learnt product\u2019s average rating embeddings (PRE) by our model. Improvements of our model are statistically significant with p<0.005 based on paired t-test.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n P || HotelP || RestaurantR || HotelR || RestaurantF1 || HotelF1 || RestaurantA || HotelA || Restaurant
Features || LF54.553.871.180.861.764.655.955.8
Features || LF+BF63.458.152.661.257.559.661.158.5
Features || BF EditSim+LF55.353.969.782.8261.665.156.656.0
Features || BF W2Vsim+W2V58.456.365.973.461.963.759.558.2
Features || Ours RE62.158.468.375.165.165.763.360.8
Features || Ours RE+RRE+PRE63.659.071.278.867.267.565.362.0
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{SVM classification results across linguistic features (LF, bigrams here (Mukherjee et al., 2013b)), behavioral features (BF: RL, RD, MCS (Mukherjee et al., 2013b)); the SVM classification results by the intuitive method that finding the most similar existing review by edit distance ratio and take the found reviewers\u2019 behavioral features as approximation (BF EditSim+LF), and results by the intuitive method that finding the most similar existing review by averaged pre-trained word embeddings (using Word2Vec) (BF W2Vsim+W2V); and the SVM classification results across the learnt review embeddings (RE), the learnt review\u2019s rating embeddings (RRE), the learnt product\u2019s average rating embeddings (PRE) by our model. Improvements of our model are statistically significant with p$<$0.005 based on paired t-test.}\n\\begin{tabular}{lrrrrrrrr}\n\\toprule\n & P $\\parallel$ Hotel & P $\\parallel$ Restaurant & R $\\parallel$ Hotel & R $\\parallel$ Restaurant & F1 $\\parallel$ Hotel & F1 $\\parallel$ Restaurant & A $\\parallel$ Hotel & A $\\parallel$ Restaurant \\\\\n\\midrule\nFeatures $\\parallel$ LF & 54.5 & 53.8 & 71.1 & 80.8 & 61.7 & 64.6 & 55.9 & 55.8 \\\\\nFeatures $\\parallel$ LF+BF & 63.4 & 58.1 & 52.6 & 61.2 & 57.5 & 59.6 & 61.1 & 58.5 \\\\\nFeatures $\\parallel$ BF EditSim+LF & 55.3 & 53.9 & 69.7 & 82.82 & 61.6 & 65.1 & 56.6 & 56 \\\\\nFeatures $\\parallel$ BF W2Vsim+W2V & 58.4 & 56.3 & 65.9 & 73.4 & 61.9 & 63.7 & 59.5 & 58.2 \\\\\nFeatures $\\parallel$ Ours RE & 62.1 & 58.4 & 68.3 & 75.1 & 65.1 & 65.7 & 63.3 & 60.8 \\\\\nFeatures $\\parallel$ Ours RE+RRE+PRE & 63.6 & 59 & 71.2 & 78.8 & 67.2 & 67.5 & 65.3 & 62 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: SVM classification results across linguistic features (LF, bigrams here (Mukherjee et al., 2013b)), behavioral features (BF: RL, RD, MCS (Mukherjee et al., 2013b)); the SVM classification results by the intuitive method that finding the most similar existing review by edit distance ratio and take the found reviewers\u2019 behavioral features as approximation (BF EditSim+LF), and results by the intuitive method that finding the most similar existing review by averaged pre-trained word embeddings (using Word2Vec) (BF W2Vsim+W2V); and the SVM classification results across the learnt review embeddings (RE), the learnt review\u2019s rating embeddings (RRE), the learnt product\u2019s average rating embeddings (PRE) by our model. Improvements of our model are statistically significant with p<0.005 based on paired t-test.\n
\n \n P || Hotel\n \n P || Restaurant\n \n R || Hotel\n \n R || Restaurant\n \n F1 || Hotel\n \n F1 || Restaurant\n \n A || Hotel\n \n A || Restaurant\n
\n Features || LF\n \n 54.5\n \n 53.8\n \n 71.1\n \n 80.8\n \n 61.7\n \n 64.6\n \n 55.9\n \n 55.8\n
\n Features || LF+BF\n \n 63.4\n \n 58.1\n \n 52.6\n \n 61.2\n \n 57.5\n \n 59.6\n \n 61.1\n \n 58.5\n
\n Features || BF EditSim+LF\n \n 55.3\n \n 53.9\n \n 69.7\n \n 82.82\n \n 61.6\n \n 65.1\n \n 56.6\n \n 56.0\n
\n Features || BF W2Vsim+W2V\n \n 58.4\n \n 56.3\n \n 65.9\n \n 73.4\n \n 61.9\n \n 63.7\n \n 59.5\n \n 58.2\n
\n Features || Ours RE\n \n 62.1\n \n 58.4\n \n 68.3\n \n 75.1\n \n 65.1\n \n 65.7\n \n 63.3\n \n 60.8\n
\n Features || Ours RE+RRE+PRE\n \n 63.6\n \n 59.0\n \n 71.2\n \n 78.8\n \n 67.2\n \n 67.5\n \n 65.3\n \n 62.0\n
\n" }, { "instance_id": "f3413007-6cb7-4e27-a996-029e4d86d9f3", "table_id_paper": "P17-1054table_2", "caption": "The performance of predicting present keyphrases of various models on five benchmark datasets", "row_header_level": 2, "row_headers": [ [ "Method", "Tf-Idf" ], [ "Method", "TextRank" ], [ "Method", "SingleRank" ], [ "Method", "ExpandRank" ], [ "Method", "Maui" ], [ "Method", "KEA" ], [ "Method", "RNN" ], [ "Method", "CopyRNN" ] ], "column_header_level": 2, "column_headers": [ [ "Inspec", "F1@5" ], [ "Inspec", "F1@10" ], [ "Krapivin", "F1@5" ], [ "Krapivin", "F1@10" ], [ "NUS", "F1@5" ], [ "NUS", "F1@10" ], [ "SemEval", "F1@5" ], [ "SemEval", "F1@10" ], [ "KP20k", "F1@5" ], [ "KP20k", "F1@10" ] ], "contents": [ [ "0.221", "0.313", "0.129", "0.160", "0.136", "0.184", "0.128", "0.194", "0.102", "0.126" ], [ "0.223", "0.281", "0.189", "0.162", "0.195", "0.196", "0.176", "0.187", "0.175", "0.147" ], [ "0.214", "0.306", "0.189", "0.162", "0.140", "0.173", "0.135", "0.176", "0.096", "0.119" ], [ "0.210", "0.304", "0.081", "0.126", "0.132", "0.164", "0.139", "0.170", "N/A", "N/A" ], [ "0.040", "0.042", "0.249", "0.216", "0.249", "0.268", "0.044", "0.039", "0.270", "0.230" ], [ "0.098", "0.126", "0.110", "0.152", "0.069", "0.084", "0.025", "0.026", "0.171", "0.154" ], [ "0.085", "0.064", "0.135", "0.088", "0.169", "0.127", "0.157", "0.124", "0.179", "0.189" ], [ "0.278", "0.342", "0.311", "0.266", "0.334", "0.326", "0.293", "0.304", "0.333", "0.262" ] ], "metrics_loc": "column", "metrics_type": "['F1@5', 'F1@10', 'F1@5', 'F1@10', 'F1@5', 'F1@10', 'F1@5', 'F1@10', 'F1@5', 'F1@10']", "target_entity": "['RNN', 'CopyRNN']", "table_html": "
Inspec || F1@5 Inspec || F1@10 Krapivin || F1@5 Krapivin || F1@10 NUS || F1@5 NUS || F1@10 SemEval || F1@5 SemEval || F1@10 KP20k || F1@5 KP20k || F1@10
Method || Tf-Idf 0.221 0.313 0.129 0.160 0.136 0.184 0.128 0.194 0.102 0.126
Method || TextRank 0.223 0.281 0.189 0.162 0.195 0.196 0.176 0.187 0.175 0.147
Method || SingleRank 0.214 0.306 0.189 0.162 0.140 0.173 0.135 0.176 0.096 0.119
Method || ExpandRank 0.210 0.304 0.081 0.126 0.132 0.164 0.139 0.170 N/A N/A
Method || Maui 0.040 0.042 0.249 0.216 0.249 0.268 0.044 0.039 0.270 0.230
Method || KEA 0.098 0.126 0.110 0.152 0.069 0.084 0.025 0.026 0.171 0.154
Method || RNN 0.085 0.064 0.135 0.088 0.169 0.127 0.157 0.124 0.179 0.189
Method || CopyRNN 0.278 0.342 0.311 0.266 0.334 0.326 0.293 0.304 0.333 0.262
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P17-1054", "page_no": 7, "dir": "acl2017", "description": "Table 2 provides the performances of the six baseline models, as well as our proposed models (i.e., RNN and CopyRNN). For each method, the table lists its F-measure at top 5 and top 10 predictions on the five datasets. The best scores are highlighted in bold and the underlines indicate the second best performances. The results show that the four unsupervised models (Tf-idf, TextRank, SingleRank and ExpandRank) have a robust performance across different datasets. The ExpandRank fails to return any result on the KP20k dataset, due to its high time complexity. The measures on NUS and SemEval here are higher than the ones reported in (Hasan and Ng, 2010) and (Kim et al., 2010), probably because we utilized the paper abstract instead of the full text for training, which may filter out some noisy information. The performance of the two supervised models (i.e., Maui and KEA) were unstable on some datasets, but Maui achieved the best performances on three datasets among all the baseline models. As for our proposed keyphrase prediction approaches, the RNN model with the attention mechanism did not perform as well as we expected. It might be because the RNN model is only concerned with finding the hidden semantics behind the text, which may tend to generate keyphrases or words that are too general and may not necessarily refer to the source text. In addition, we observe that 2.5% (70,891/2,780,316) of keyphrases in our dataset contain out-of-vocabulary words, which the RNN model is not able to recall, since the RNN model can only generate results with the 50,000 words in vocabulary. This indicates that a pure generative model may not fit the extraction task, and we need to further link back to the language usage within the source text. The CopyRNN model, by considering more contextual information, significantly outperforms not only the RNN model but also all baselines, exceeding the best baselines by more than 20% on average.", "class_sentence": "[1, 1, 2, 1, 2, 2, 1, 1, 2, 2, 2, 1]", "sentences": "['Table 2 provides the performances of the six baseline models, as well as our proposed models (i.e., RNN and CopyRNN).', 'For each method, the table lists its F-measure at top 5 and top 10 predictions on the five datasets.', 'The best scores are highlighted in bold and the underlines indicate the second best performances.', 'The results show that the four unsupervised models (Tf-idf, TextRank, SingleRank and ExpandRank) have a robust performance across different datasets.', 'The ExpandRank fails to return any result on the KP20k dataset, due to its high time complexity.', 'The measures on NUS and SemEval here are higher than the ones reported in (Hasan and Ng, 2010) and (Kim et al., 2010), probably because we utilized the paper abstract instead of the full text for training, which may filter out some noisy information.', 'The performance of the two supervised models (i.e., Maui and KEA) were unstable on some datasets, but Maui achieved the best performances on three datasets among all the baseline models.', 'As for our proposed keyphrase prediction approaches, the RNN model with the attention mechanism did not perform as well as we expected.', 'It might be because the RNN model is only concerned with finding the hidden semantics behind the text, which may tend to generate keyphrases or words that are too general and may not necessarily refer to the source text.', 'In addition, we observe that 2.5% (70,891/2,780,316) of keyphrases in our dataset contain out-of-vocabulary words, which the RNN model is not able to recall, since the RNN model can only generate results with the 50,000 words in vocabulary.', 'This indicates that a pure generative model may not fit the extraction task, and we need to further link back to the language usage within the source text.', 'The CopyRNN model, by considering more contextual information, significantly outperforms not only the RNN model but also all baselines, exceeding the best baselines by more than 20% on average.']", "header_mention": "[['RNN', 'CopyRNN'], ['F1@5', 'F1@10', 'Inspec', 'Krapivin', 'NUS', 'SemEval', 'KP20k'], None, ['Tf-Idf', 'TextRank', 'SingleRank', 'ExpandRank'], ['ExpandRank', 'KP20k'], ['NUS', 'SemEval'], ['Maui', 'KEA'], ['RNN'], ['RNN'], ['RNN'], ['RNN'], ['CopyRNN', 'RNN']]", "valid": 1, "image_id": "P17-1054-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

The performance of predicting present keyphrases of various models on five benchmark datasets

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Inspec || F1@5Inspec || F1@10Krapivin || F1@5Krapivin || F1@10NUS || F1@5NUS || F1@10SemEval || F1@5SemEval || F1@10KP20k || F1@5KP20k || F1@10
Method || Tf-Idf0.2210.3130.1290.1600.1360.1840.1280.1940.1020.126
Method || TextRank0.2230.2810.1890.1620.1950.1960.1760.1870.1750.147
Method || SingleRank0.2140.3060.1890.1620.1400.1730.1350.1760.0960.119
Method || ExpandRank0.2100.3040.0810.1260.1320.1640.1390.170N/AN/A
Method || Maui0.0400.0420.2490.2160.2490.2680.0440.0390.2700.230
Method || KEA0.0980.1260.1100.1520.0690.0840.0250.0260.1710.154
Method || RNN0.0850.0640.1350.0880.1690.1270.1570.1240.1790.189
Method || CopyRNN0.2780.3420.3110.2660.3340.3260.2930.3040.3330.262
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{The performance of predicting present keyphrases of various models on five benchmark datasets}\n\\begin{tabular}{lrrrrrrrrll}\n\\toprule\n & Inspec $\\parallel$ F1@5 & Inspec $\\parallel$ F1@10 & Krapivin $\\parallel$ F1@5 & Krapivin $\\parallel$ F1@10 & NUS $\\parallel$ F1@5 & NUS $\\parallel$ F1@10 & SemEval $\\parallel$ F1@5 & SemEval $\\parallel$ F1@10 & KP20k $\\parallel$ F1@5 & KP20k $\\parallel$ F1@10 \\\\\n\\midrule\nMethod $\\parallel$ Tf-Idf & 0.221 & 0.313 & 0.129 & 0.16 & 0.136 & 0.184 & 0.128 & 0.194 & 0.102 & 0.126 \\\\\nMethod $\\parallel$ TextRank & 0.223 & 0.281 & 0.189 & 0.162 & 0.195 & 0.196 & 0.176 & 0.187 & 0.175 & 0.147 \\\\\nMethod $\\parallel$ SingleRank & 0.214 & 0.306 & 0.189 & 0.162 & 0.14 & 0.173 & 0.135 & 0.176 & 0.096 & 0.119 \\\\\nMethod $\\parallel$ ExpandRank & 0.21 & 0.304 & 0.081 & 0.126 & 0.132 & 0.164 & 0.139 & 0.17 & & \\\\\nMethod $\\parallel$ Maui & 0.04 & 0.042 & 0.249 & 0.216 & 0.249 & 0.268 & 0.044 & 0.039 & 0.27 & 0.23 \\\\\nMethod $\\parallel$ KEA & 0.098 & 0.126 & 0.11 & 0.152 & 0.069 & 0.084 & 0.025 & 0.026 & 0.171 & 0.154 \\\\\nMethod $\\parallel$ RNN & 0.085 & 0.064 & 0.135 & 0.088 & 0.169 & 0.127 & 0.157 & 0.124 & 0.179 & 0.189 \\\\\nMethod $\\parallel$ CopyRNN & 0.278 & 0.342 & 0.311 & 0.266 & 0.334 & 0.326 & 0.293 & 0.304 & 0.333 & 0.262 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: The performance of predicting present keyphrases of various models on five benchmark datasets\n
\n \n Inspec || F1@5\n \n Inspec || F1@10\n \n Krapivin || F1@5\n \n Krapivin || F1@10\n \n NUS || F1@5\n \n NUS || F1@10\n \n SemEval || F1@5\n \n SemEval || F1@10\n \n KP20k || F1@5\n \n KP20k || F1@10\n
\n Method || Tf-Idf\n \n 0.221\n \n 0.313\n \n 0.129\n \n 0.160\n \n 0.136\n \n 0.184\n \n 0.128\n \n 0.194\n \n 0.102\n \n 0.126\n
\n Method || TextRank\n \n 0.223\n \n 0.281\n \n 0.189\n \n 0.162\n \n 0.195\n \n 0.196\n \n 0.176\n \n 0.187\n \n 0.175\n \n 0.147\n
\n Method || SingleRank\n \n 0.214\n \n 0.306\n \n 0.189\n \n 0.162\n \n 0.140\n \n 0.173\n \n 0.135\n \n 0.176\n \n 0.096\n \n 0.119\n
\n Method || ExpandRank\n \n 0.210\n \n 0.304\n \n 0.081\n \n 0.126\n \n 0.132\n \n 0.164\n \n 0.139\n \n 0.170\n \n N/A\n \n N/A\n
\n Method || Maui\n \n 0.040\n \n 0.042\n \n 0.249\n \n 0.216\n \n 0.249\n \n 0.268\n \n 0.044\n \n 0.039\n \n 0.270\n \n 0.230\n
\n Method || KEA\n \n 0.098\n \n 0.126\n \n 0.110\n \n 0.152\n \n 0.069\n \n 0.084\n \n 0.025\n \n 0.026\n \n 0.171\n \n 0.154\n
\n Method || RNN\n \n 0.085\n \n 0.064\n \n 0.135\n \n 0.088\n \n 0.169\n \n 0.127\n \n 0.157\n \n 0.124\n \n 0.179\n \n 0.189\n
\n Method || CopyRNN\n \n 0.278\n \n 0.342\n \n 0.311\n \n 0.266\n \n 0.334\n \n 0.326\n \n 0.293\n \n 0.304\n \n 0.333\n \n 0.262\n
\n" }, { "instance_id": "eacc8677-a017-43eb-bd33-dd7e619d1599", "table_id_paper": "P17-1176table_4", "caption": "Comparison of our proposed methods on Spanish-French and German-French translation tasks from the Europarl corpus. English is treated as the pivot language.", "row_header_level": 2, "row_headers": [ [ "Method", "sent-greedy" ], [ "Method", "sent-beam" ], [ "Method", "word-greedy" ], [ "Method", "word-beam" ], [ "Method", "word-sampling" ] ], "column_header_level": 2, "column_headers": [ [ "Es\u2192 Fr", "dev" ], [ "Es\u2192 Fr", "test" ], [ "De\u2192 Fr", "dev" ], [ "De\u2192 Fr", "test" ] ], "contents": [ [ "31.00", "31.05", "22.34", "21.88" ], [ "31.57", "31.64", "24.95", "24.39" ], [ "31.37", "31.92", "24.72", "25.15" ], [ "30.81", "31.21", "24.64", "24.19" ], [ "33.65", "33.86", "26.99", "27.03" ] ], "metrics_loc": "column", "metrics_type": "['BLEU', 'BLEU', 'BLEU', 'BLEU']", "target_entity": "['Method']", "table_html": "
Es\u2192 Fr || dev Es\u2192 Fr || test De\u2192 Fr || dev De\u2192 Fr || test
Method || sent-greedy 31.00 31.05 22.34 21.88
Method || sent-beam 31.57 31.64 24.95 24.39
Method || word-greedy 31.37 31.92 24.72 25.15
Method || word-beam 30.81 31.21 24.64 24.19
Method || word-sampling 33.65 33.86 26.99 27.03
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P17-1176", "page_no": 6, "dir": "acl2017", "description": "Table 4 shows BLEU scores on the Europarl corpus of our proposed methods. For sentence-level approaches, the sent-beam method outperforms the sent-greedy method by +0.59 BLEU points over Spanish-French translation and +2.51 BLEU points over German-French translation on the test set. For word-level experiments, we observe that the word-sampling method performs much better than the other two methods: +1.94 BLEU points on Spanish-French translation and +1.88 BLEU points on German-French translation over the word-greedy method; +2.65 BLEU points on Spanish-French translation and +2.84 BLEU points on German-French translation over the word-beam method.", "class_sentence": "[1, 1, 1]", "sentences": "['Table 4 shows BLEU scores on the Europarl corpus of our proposed methods.', 'For sentence-level approaches, the sent-beam method outperforms the sent-greedy method by +0.59 BLEU points over Spanish-French translation and +2.51 BLEU points over German-French translation on the test set.', 'For word-level experiments, we observe that the word-sampling method performs much better than the other two methods: +1.94 BLEU points on Spanish-French translation and +1.88 BLEU points on German-French translation over the word-greedy method; +2.65 BLEU points on Spanish-French translation and +2.84 BLEU points on German-French translation over the word-beam method.']", "header_mention": "[None, ['sent-beam', 'sent-greedy', 'Es\u2192 Fr', 'test'], ['word-sampling', 'Es\u2192 Fr', 'De\u2192 Fr', 'word-greedy', 'word-beam']]", "valid": 1, "image_id": "P17-1176-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison of our proposed methods on Spanish-French and German-French translation tasks from the Europarl corpus. English is treated as the pivot language.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Es\u2192 Fr || devEs\u2192 Fr || testDe\u2192 Fr || devDe\u2192 Fr || test
Method || sent-greedy31.0031.0522.3421.88
Method || sent-beam31.5731.6424.9524.39
Method || word-greedy31.3731.9224.7225.15
Method || word-beam30.8131.2124.6424.19
Method || word-sampling33.6533.8626.9927.03
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison of our proposed methods on Spanish-French and German-French translation tasks from the Europarl corpus. English is treated as the pivot language.}\n\\begin{tabular}{lrrrr}\n\\toprule\n & Es$\\rightarrow$ Fr $\\parallel$ dev & Es$\\rightarrow$ Fr $\\parallel$ test & De$\\rightarrow$ Fr $\\parallel$ dev & De$\\rightarrow$ Fr $\\parallel$ test \\\\\n\\midrule\nMethod $\\parallel$ sent-greedy & 31 & 31.05 & 22.34 & 21.88 \\\\\nMethod $\\parallel$ sent-beam & 31.57 & 31.64 & 24.95 & 24.39 \\\\\nMethod $\\parallel$ word-greedy & 31.37 & 31.92 & 24.72 & 25.15 \\\\\nMethod $\\parallel$ word-beam & 30.81 & 31.21 & 24.64 & 24.19 \\\\\nMethod $\\parallel$ word-sampling & 33.65 & 33.86 & 26.99 & 27.03 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Comparison of our proposed methods on Spanish-French and German-French translation tasks from the Europarl corpus. English is treated as the pivot language.\n
\n \n Es\u2192 Fr || dev\n \n Es\u2192 Fr || test\n \n De\u2192 Fr || dev\n \n De\u2192 Fr || test\n
\n Method || sent-greedy\n \n 31.00\n \n 31.05\n \n 22.34\n \n 21.88\n
\n Method || sent-beam\n \n 31.57\n \n 31.64\n \n 24.95\n \n 24.39\n
\n Method || word-greedy\n \n 31.37\n \n 31.92\n \n 24.72\n \n 25.15\n
\n Method || word-beam\n \n 30.81\n \n 31.21\n \n 24.64\n \n 24.19\n
\n Method || word-sampling\n \n 33.65\n \n 33.86\n \n 26.99\n \n 27.03\n
\n" }, { "instance_id": "12afd732-757d-4d28-b013-5fa28d37043b", "table_id_paper": "P17-1183table_1", "caption": "Results on the CELEX dataset", "row_header_level": 1, "row_headers": [ [ "MED (Kann and Schutze 2016a)" ], [ "NWFST (Rastogi et al. 2016)" ], [ "LAT (Dreyer et al. 2008)" ], [ "Soft" ], [ "Hard" ] ], "column_header_level": 1, "column_headers": [ [ "13SIA" ], [ "2PIE" ], [ "2PKE" ], [ "rP" ], [ "Avg." ] ], "contents": [ [ "83.9", "95", "87.6", "84", "87.62" ], [ "86.8", "94.8", "87.9", "81.1", "87.65" ], [ "87.5", "93.4", "87.4", "84.9", "88.3" ], [ "83.1", "93.8", "88", "83.2", "87" ], [ "85.8", "95.1", "89.5", "87.2", "89.44" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy', 'accuracy', 'avg accuracy']", "target_entity": "['Hard', 'Soft']", "table_html": "
13SIA 2PIE 2PKE rP Avg.
MED (Kann and Schutze 2016a) 83.9 95 87.6 84 87.62
NWFST (Rastogi et al. 2016) 86.8 94.8 87.9 81.1 87.65
LAT (Dreyer et al. 2008) 87.5 93.4 87.4 84.9 88.3
Soft 83.1 93.8 88 83.2 87
Hard 85.8 95.1 89.5 87.2 89.44
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P17-1183", "page_no": 6, "dir": "acl2017", "description": "On the low resource setting (CELEX), our hard attention model significantly outperforms both the recent neural models of Kann and Schutze (2016a) (MED) and Rastogi et al. (2016) (NWFST) and the morphologically aware latent variable model of Dreyer et al. (2008) (LAT), as detailed in Table 1. In addition, it significantly outperforms our implementation of the soft attention model (Soft). It is also, to our knowledge, the first model that surpassed in overall accuracy the latent variable model on this dataset. We attribute our advantage over the soft attention models to the ability of the hard attention control mechanism to harness the monotonic alignments found in the data. The advantage over the FST models may be explained by our conditioning on the entire output history which is not available in those models.", "class_sentence": "[1, 1, 2, 2, 2]", "sentences": "['On the low resource setting (CELEX), our hard attention model significantly outperforms both the recent neural models of Kann and Schutze (2016a) (MED) and Rastogi et al. (2016) (NWFST) and the morphologically aware latent variable model of Dreyer et al. (2008) (LAT), as detailed in Table 1.', 'In addition, it significantly outperforms our implementation of the soft attention model (Soft).', 'It is also, to our knowledge, the first model that surpassed in overall accuracy the latent variable model on this dataset.', 'We attribute our advantage over the soft attention models to the ability of the hard attention control mechanism to harness the monotonic alignments found in the data.', 'The advantage over the FST models may be explained by our conditioning on the entire output history which is not available in those models.']", "header_mention": "[['Hard', 'MED (Kann and Schutze 2016a)', 'NWFST (Rastogi et al. 2016)', 'LAT (Dreyer et al. 2008)'], ['Hard', 'Soft'], ['Hard'], ['Soft', 'Hard'], ['NWFST (Rastogi et al. 2016)']]", "valid": 1, "image_id": "P17-1183-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results on the CELEX dataset

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n 13SIA2PIE2PKErPAvg.
MED (Kann and Schutze 2016a)83.99587.68487.62
NWFST (Rastogi et al. 2016)86.894.887.981.187.65
LAT (Dreyer et al. 2008)87.593.487.484.988.3
Soft83.193.88883.287
Hard85.895.189.587.289.44
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results on the CELEX dataset}\n\\begin{tabular}{lrrrrr}\n\\toprule\n & 13SIA & 2PIE & 2PKE & rP & Avg. \\\\\n\\midrule\nMED (Kann and Schutze 2016a) & 83.9 & 95 & 87.6 & 84 & 87.62 \\\\\nNWFST (Rastogi et al. 2016) & 86.8 & 94.8 & 87.9 & 81.1 & 87.65 \\\\\nLAT (Dreyer et al. 2008) & 87.5 & 93.4 & 87.4 & 84.9 & 88.3 \\\\\nSoft & 83.1 & 93.8 & 88 & 83.2 & 87 \\\\\nHard & 85.8 & 95.1 & 89.5 & 87.2 & 89.44 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Results on the CELEX dataset\n
\n \n 13SIA\n \n 2PIE\n \n 2PKE\n \n rP\n \n Avg.\n
\n MED (Kann and Schutze 2016a)\n \n 83.9\n \n 95\n \n 87.6\n \n 84\n \n 87.62\n
\n NWFST (Rastogi et al. 2016)\n \n 86.8\n \n 94.8\n \n 87.9\n \n 81.1\n \n 87.65\n
\n LAT (Dreyer et al. 2008)\n \n 87.5\n \n 93.4\n \n 87.4\n \n 84.9\n \n 88.3\n
\n Soft\n \n 83.1\n \n 93.8\n \n 88\n \n 83.2\n \n 87\n
\n Hard\n \n 85.8\n \n 95.1\n \n 89.5\n \n 87.2\n \n 89.44\n
\n" }, { "instance_id": "5293d80d-3732-49fb-9fd5-cc7d0392d0f7", "table_id_paper": "P17-1195table_1", "caption": "Performance of the reasoning module on manually formalized pre-university problems", "row_header_level": 2, "row_headers": [ [ "Dataset", "DEV" ], [ "Dataset", "TEST" ] ], "column_header_level": 2, "column_headers": [ [ "Succeeded", "Success %" ], [ "Succeeded", "Avg. Time" ], [ "Failed", "Timeout" ], [ "Failed", "Other" ] ], "contents": [ [ "75.3% (131/174)", "10.5s", "16.70%", "8.10%" ], [ "78.2% (172/220)", "16.2s", "15.00%", "6.80%" ] ], "metrics_loc": "column", "metrics_type": "['Success %', 'Avg. Time', 'Timeout', 'Other']", "target_entity": "['Succeeded']", "table_html": "
Succeeded || Success % Succeeded || Avg. Time Failed || Timeout Failed || Other
Dataset || DEV 75.3% (131/174) 10.5s 16.70% 8.10%
Dataset || TEST 78.2% (172/220) 16.2s 15.00% 6.80%
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P17-1195", "page_no": 3, "dir": "acl2017", "description": "The reasoning module (i.e., the formula rewriting and the deduction with CAS and ATP) of the system has been extensively tested on a large collection of manually formalized pre-university math problems that includes more than 1,500 problems. It solves 70% of the them in the time limit of 10 minutes per problem. Table 1 shows the rate of successfully solved problems in the manually formalized version of the benchmark problems used in the current paper.", "class_sentence": "[2, 1, 1]", "sentences": "['The reasoning module (i.e., the formula rewriting and the deduction with CAS and ATP) of the system has been extensively tested on a large collection of manually formalized pre-university math problems that includes more than 1,500 problems.', 'It solves 70% of the them in the time limit of 10 minutes per problem.', 'Table 1 shows the rate of successfully solved problems in the manually formalized version of the benchmark problems used in the current paper.']", "header_mention": "[None, ['Success %', 'Avg. Time'], ['Success %']]", "valid": 1, "image_id": "P17-1195-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance of the reasoning module on manually formalized pre-university problems

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Succeeded || Success %Succeeded || Avg. TimeFailed || TimeoutFailed || Other
Dataset || DEV75.3% (131/174)10.5s16.70%8.10%
Dataset || TEST78.2% (172/220)16.2s15.00%6.80%
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance of the reasoning module on manually formalized pre-university problems}\n\\begin{tabular}{lllll}\n\\toprule\n & Succeeded $\\parallel$ Success \\% & Succeeded $\\parallel$ Avg. Time & Failed $\\parallel$ Timeout & Failed $\\parallel$ Other \\\\\n\\midrule\nDataset $\\parallel$ DEV & 75.3\\% (131/174) & 10.5s & 16.70\\% & 8.10\\% \\\\\nDataset $\\parallel$ TEST & 78.2\\% (172/220) & 16.2s & 15.00\\% & 6.80\\% \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Performance of the reasoning module on manually formalized pre-university problems\n
\n \n Succeeded || Success %\n \n Succeeded || Avg. Time\n \n Failed || Timeout\n \n Failed || Other\n
\n Dataset || DEV\n \n 75.3% (131/174)\n \n 10.5s\n \n 16.70%\n \n 8.10%\n
\n Dataset || TEST\n \n 78.2% (172/220)\n \n 16.2s\n \n 15.00%\n \n 6.80%\n
\n" }, { "instance_id": "b5bb7e6c-d878-41ab-9dd7-29316c171129", "table_id_paper": "P17-2030table_1", "caption": "Unlabeled dependency accuracy results with the 5x5 models and test sets. \u2207 shows the slope of deterioration in parser performance.", "row_header_level": 1, "row_headers": [ [ "0%" ], [ "5%" ], [ "10%" ], [ "15%" ], [ "20%" ] ], "column_header_level": 1, "column_headers": [ [ "Baseline" ], [ "E05" ], [ "E10" ], [ "E15" ], [ "E20" ] ], "contents": [ [ "91.43", "91.12", "90.87", "90.61", "90.29" ], [ "89.99", "90", "89.87", "89.72", "89.48" ], [ "87.84", "87.99", "88.07", "88.14", "88.04" ], [ "85.64", "86.18", "86.54", "86.75", "86.82" ], [ "84.12", "84.78", "85.28", "85.5", "85.76" ] ], "metrics_loc": "row", "metrics_type": "['UAS', 'UAS', 'UAS', 'UAS', 'UAS']", "target_entity": "['0%', '5%', '10%', '15%', '20%']", "table_html": "
Baseline E05 E10 E15 E20
0% 91.43 91.12 90.87 90.61 90.29
5% 89.99 90 89.87 89.72 89.48
10% 87.84 87.99 88.07 88.14 88.04
15% 85.64 86.18 86.54 86.75 86.82
20% 84.12 84.78 85.28 85.5 85.76
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P17-2030", "page_no": 4, "dir": "acl2017", "description": "Table 1 shows the result of unlabeled dependency accuracy (UAS). On the error-free test set (0%), the baseline (EF pipeline) outperforms other EREF models; the accuracy is lower when the parser is trained on noisier data. The difference among the models becomes small when the test set has 10% error injection rate. As the rate increases further, the trend of parser accuracy reverses. When the test set has 15% or higher noise, the E20 is the most accurate parser.", "class_sentence": "[1, 1, 1, 1, 1]", "sentences": "['Table 1 shows the result of unlabeled dependency accuracy (UAS).', 'On the error-free test set (0%), the baseline (EF pipeline) outperforms other EREF models; the accuracy is lower when the parser is trained on noisier data.', 'The difference among the models becomes small when the test set has 10% error injection rate.', 'As the rate increases further, the trend of parser accuracy reverses.', 'When the test set has 15% or higher noise, the E20 is the most accurate parser.']", "header_mention": "[None, ['0%', 'Baseline', 'E05', 'E10', 'E15', 'E20'], ['10%'], None, ['15%', 'E20']]", "valid": 1, "image_id": "P17-2030-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Unlabeled dependency accuracy results with the 5x5 models and test sets. \u2207 shows the slope of deterioration in parser performance.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n BaselineE05E10E15E20
0%91.4391.1290.8790.6190.29
5%89.999089.8789.7289.48
10%87.8487.9988.0788.1488.04
15%85.6486.1886.5486.7586.82
20%84.1284.7885.2885.585.76
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Unlabeled dependency accuracy results with the 5x5 models and test sets. $\\nabla$ shows the slope of deterioration in parser performance.}\n\\begin{tabular}{lrrrrr}\n\\toprule\n & Baseline & E05 & E10 & E15 & E20 \\\\\n\\midrule\n0\\% & 91.43 & 91.12 & 90.87 & 90.61 & 90.29 \\\\\n5\\% & 89.99 & 90 & 89.87 & 89.72 & 89.48 \\\\\n10\\% & 87.84 & 87.99 & 88.07 & 88.14 & 88.04 \\\\\n15\\% & 85.64 & 86.18 & 86.54 & 86.75 & 86.82 \\\\\n20\\% & 84.12 & 84.78 & 85.28 & 85.5 & 85.76 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Unlabeled dependency accuracy results with the 5x5 models and test sets. \u2207 shows the slope of deterioration in parser performance.\n
\n \n Baseline\n \n E05\n \n E10\n \n E15\n \n E20\n
\n 0%\n \n 91.43\n \n 91.12\n \n 90.87\n \n 90.61\n \n 90.29\n
\n 5%\n \n 89.99\n \n 90\n \n 89.87\n \n 89.72\n \n 89.48\n
\n 10%\n \n 87.84\n \n 87.99\n \n 88.07\n \n 88.14\n \n 88.04\n
\n 15%\n \n 85.64\n \n 86.18\n \n 86.54\n \n 86.75\n \n 86.82\n
\n 20%\n \n 84.12\n \n 84.78\n \n 85.28\n \n 85.5\n \n 85.76\n
\n" }, { "instance_id": "d401c3f0-7cb1-4c78-af66-a86000c94a42", "table_id_paper": "P18-1008table_3", "caption": "Performance comparison. Examples/s are normalized by the number of GPUs used in the training job. FLOPs are computed assuming that source and target sequence length are both 50.", "row_header_level": 2, "row_headers": [ [ "Model", "ConvS2S" ], [ "Model", "Trans. Base" ], [ "Model", "Trans. Big" ], [ "Model", "RNMT+" ] ], "column_header_level": 1, "column_headers": [ [ "Examples/s" ], [ "FLOPs" ], [ "Params" ] ], "contents": [ [ "80", "15.7B", "263.4M" ], [ "160", "6.2B", "93.3M" ], [ "50", "31.2B", "375.4M" ], [ "30", "28.1B", "378.9M" ] ], "metrics_loc": "column", "metrics_type": "['Example/s', 'FLOPs', 'Params']", "target_entity": "['RNMT+']", "table_html": "
Examples/s FLOPs Params
Model || ConvS2S 80 15.7B 263.4M
Model || Trans. Base 160 6.2B 93.3M
Model || Trans. Big 50 31.2B 375.4M
Model || RNMT+ 30 28.1B 378.9M
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "P18-1008", "page_no": 6, "dir": "acl2018", "description": "Table 3 summarizes training performance and model statistics. The Transformer Base model is the fastest model in terms of training speed. RNMT+ is slower to train than the Transformer\nBig model on a per-GPU basis. However, sinc the RNMT+ model is quite stable, we were able to offset the lower per-GPU throughput with higher\nconcurrency by increasing the number of model replicas, and hence the overall time to convergence was not slowed down much. We also computed the number of floating point operations (FLOPs) in the model\u0081fs forward path as well as the number of total parameters for all architectures (cf. Table 3). RNMT+ requires fewer FLOPs than the Transformer Big model, even though both models have a comparable number of parameters.", "class_sentence": "[1, 1, 1, 1, 1, 1]", "sentences": "['Table 3 summarizes training performance and model statistics.', 'The Transformer Base model is the fastest model in terms of training speed.', 'RNMT+ is slower to train than the Transformer\\nBig model on a per-GPU basis.', 'However, sinc the RNMT+ model is quite stable, we were able to offset the lower per-GPU throughput with higher\\nconcurrency by increasing the number of model replicas, and hence the overall time to convergence was not slowed down much.', 'We also computed the number of floating point operations (FLOPs) in the model\\x81fs forward path as well as the number of total parameters for all architectures (cf. Table 3).', 'RNMT+ requires fewer FLOPs than the Transformer Big model, even though both models have a comparable number of parameters.']", "header_mention": "[None, ['Trans. Base'], ['RNMT+', 'Trans. Base'], ['RNMT+'], ['FLOPs', 'ConvS2S', 'Trans. Base', 'Trans. Big', 'RNMT+'], ['RNMT+', 'Trans. Big']]", "valid": 1, "image_id": "P18-1008-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance comparison. Examples/s are normalized by the number of GPUs used in the training job. FLOPs are computed assuming that source and target sequence length are both 50.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Examples/sFLOPsParams
Model || ConvS2S8015.7B263.4M
Model || Trans. Base1606.2B93.3M
Model || Trans. Big5031.2B375.4M
Model || RNMT+3028.1B378.9M
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance comparison. Examples/s are normalized by the number of GPUs used in the training job. FLOPs are computed assuming that source and target sequence length are both 50.}\n\\begin{tabular}{lrll}\n\\toprule\n & Examples/s & FLOPs & Params \\\\\n\\midrule\nModel $\\parallel$ ConvS2S & 80 & 15.7B & 263.4M \\\\\nModel $\\parallel$ Trans. Base & 160 & 6.2B & 93.3M \\\\\nModel $\\parallel$ Trans. Big & 50 & 31.2B & 375.4M \\\\\nModel $\\parallel$ RNMT+ & 30 & 28.1B & 378.9M \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Performance comparison. Examples/s are normalized by the number of GPUs used in the training job. FLOPs are computed assuming that source and target sequence length are both 50.\n
\n \n Examples/s\n \n FLOPs\n \n Params\n
\n Model || ConvS2S\n \n 80\n \n 15.7B\n \n 263.4M\n
\n Model || Trans. Base\n \n 160\n \n 6.2B\n \n 93.3M\n
\n Model || Trans. Big\n \n 50\n \n 31.2B\n \n 375.4M\n
\n Model || RNMT+\n \n 30\n \n 28.1B\n \n 378.9M\n
\n" }, { "instance_id": "7b1df1bb-9b86-473f-b7ac-d87f0c1e10aa", "table_id_paper": "P18-1021table_5", "caption": "Human evaluation results on grammaticality (Gram), informativeness (Info), and relevance (Rel) of arguments. Our model with separate decoder and attention over keyphrases receives significantly better ratings in informativeness and relevance than seq2seq (one-way ANOVA, p < 0.005).", "row_header_level": 2, "row_headers": [ [ "System", "RETRIEVAL" ], [ "System", "SEQ2SEQ" ], [ "System", "OUR MODEL" ] ], "column_header_level": 1, "column_headers": [ [ "Gram" ], [ " Info" ], [ "Rel" ] ], "contents": [ [ "4.5 \u0081} 0.6", "3.7 \u0081} 0.9", "3.3 \u0081} 1.1" ], [ "3.3 \u0081} 1.1", "1.2 \u0081} 0.5", "1.4 \u0081} 0.7" ], [ "2.5 \u0081} 0.8", "1.6 \u0081} 0.8", "1.8 \u0081} 0.8" ] ], "metrics_loc": "column", "metrics_type": "['Gram', 'Info', 'Rel']", "target_entity": "['OUR MODEL']", "table_html": "
Gram Info Rel
System || RETRIEVAL 4.5 \u0081} 0.6 3.7 \u0081} 0.9 3.3 \u0081} 1.1
System || SEQ2SEQ 3.3 \u0081} 1.1 1.2 \u0081} 0.5 1.4 \u0081} 0.7
System || OUR MODEL 2.5 \u0081} 0.8 1.6 \u0081} 0.8 1.8 \u0081} 0.8
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "P18-1021", "page_no": 8, "dir": "acl2018", "description": "Table 5 shows that our model with separate decoder and attention over keyphrases produce significantly more informative and relevant arguments than seq2seq trained without evidence.8 . However, we also observe that human judges prefer the retrieved arguments over generation-based models, illustrating the gap between system arguments and human edited text.", "class_sentence": "[1, 1]", "sentences": "['Table 5 shows that our model with separate decoder and attention over keyphrases produce significantly more informative and relevant arguments than seq2seq trained without evidence.8 .', 'However, we also observe that human judges prefer the retrieved arguments over generation-based models, illustrating the gap between system arguments and human edited text.']", "header_mention": "[['OUR MODEL', 'SEQ2SEQ'], ['RETRIEVAL', 'SEQ2SEQ', 'OUR MODEL']]", "valid": 1, "image_id": "P18-1021-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Human evaluation results on grammaticality (Gram), informativeness (Info), and relevance (Rel) of arguments. Our model with separate decoder and attention over keyphrases receives significantly better ratings in informativeness and relevance than seq2seq (one-way ANOVA, p < 0.005).

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n GramInfoRel
System || RETRIEVAL4.5 \u0081} 0.63.7 \u0081} 0.93.3 \u0081} 1.1
System || SEQ2SEQ3.3 \u0081} 1.11.2 \u0081} 0.51.4 \u0081} 0.7
System || OUR MODEL2.5 \u0081} 0.81.6 \u0081} 0.81.8 \u0081} 0.8
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Human evaluation results on grammaticality (Gram), informativeness (Info), and relevance (Rel) of arguments. Our model with separate decoder and attention over keyphrases receives significantly better ratings in informativeness and relevance than seq2seq (one-way ANOVA, p $<$ 0.005).}\n\\begin{tabular}{llll}\n\\toprule\n & Gram & Info & Rel \\\\\n\\midrule\nSystem $\\parallel$ RETRIEVAL & 4.5 $\\pm$ 0.6 & 3.7 $\\pm$ 0.9 & 3.3 $\\pm$ 1.1 \\\\\nSystem $\\parallel$ SEQ2SEQ & 3.3 $\\pm$ 1.1 & 1.2 $\\pm$ 0.5 & 1.4 $\\pm$ 0.7 \\\\\nSystem $\\parallel$ OUR MODEL & 2.5 $\\pm$ 0.8 & 1.6 $\\pm$ 0.8 & 1.8 $\\pm$ 0.8 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Human evaluation results on grammaticality (Gram), informativeness (Info), and relevance (Rel) of arguments. Our model with separate decoder and attention over keyphrases receives significantly better ratings in informativeness and relevance than seq2seq (one-way ANOVA, p < 0.005).\n
\n \n Gram\n \n Info\n \n Rel\n
\n System || RETRIEVAL\n \n 4.5 \u0081} 0.6\n \n 3.7 \u0081} 0.9\n \n 3.3 \u0081} 1.1\n
\n System || SEQ2SEQ\n \n 3.3 \u0081} 1.1\n \n 1.2 \u0081} 0.5\n \n 1.4 \u0081} 0.7\n
\n System || OUR MODEL\n \n 2.5 \u0081} 0.8\n \n 1.6 \u0081} 0.8\n \n 1.8 \u0081} 0.8\n
\n" }, { "instance_id": "f981392b-737a-49ff-b378-daca48bd0190", "table_id_paper": "P18-1038table_1", "caption": "Parsing accuracy of the best existing grammar-free and -based models as well as our SHRG-based model. Results are copied from (Oepen et al., 2015; Peng et al., 2017a; Buys and Blunsom, 2017).", "row_header_level": 4, "row_headers": [ [ "Model", "Data-driven", "Grammar", "NO" ], [ "Model", "ERG-based", "Grammar", "Unification" ], [ "Model", "SHRG-based", "Grammar", "Rewriting" ] ], "column_header_level": 1, "column_headers": [ [ "SDG" ], [ "EDS" ], [ "DMRS" ] ], "contents": [ [ "89.4", "85.48", "84.16" ], [ "92.8", "89.58", "89.64" ], [ "-", "90.39", "89.51" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy', 'accuracy']", "target_entity": "['SHRG-based']", "table_html": "
SDG EDS DMRS
Model || Data-driven || Grammar || NO 89.4 85.48 84.16
Model || ERG-based || Grammar || Unification 92.8 89.58 89.64
Model || SHRG-based || Grammar || Rewriting - 90.39 89.51
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P18-1038", "page_no": 3, "dir": "acl2018", "description": "Our parser achieves an accuracy of 90.35 for EDS and 89.51 for DMRS in terms of ELEMENTARY DEPENDENCY MATCH (EDM) which outperforms the best existing grammar-free model (Buys and Blunsom, 2017) by a significant margin (see Table 1). This marked result affirms the value of modeling the syntacto-semantic composition process for semantic parsing. On sentences that can be parsed by ERG-guided parsers, e.g. PET2 or ACE3 , significant accuracy gaps between ERG-guided parsers and data-driven parsers are repeatedly reported (see Table 1).", "class_sentence": "[1, 1, 1]", "sentences": "['Our parser achieves an accuracy of 90.35 for EDS and 89.51 for DMRS in terms of ELEMENTARY DEPENDENCY MATCH (EDM) which outperforms the best existing grammar-free model (Buys and Blunsom, 2017) by a significant margin (see Table 1).', 'This marked result affirms the value of modeling the syntacto-semantic composition process for semantic parsing.', 'On sentences that can be parsed by ERG-guided parsers, e.g. PET2 or ACE3 , significant accuracy gaps between ERG-guided parsers and data-driven parsers are repeatedly reported (see Table 1).']", "header_mention": "[['SHRG-based', 'EDS', 'DMRS'], None, ['ERG-based', 'Data-driven']]", "valid": 1, "image_id": "P18-1038-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Parsing accuracy of the best existing grammar-free and -based models as well as our SHRG-based model. Results are copied from (Oepen et al., 2015; Peng et al., 2017a; Buys and Blunsom, 2017).

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n SDGEDSDMRS
Model || Data-driven || Grammar || NO89.485.4884.16
Model || ERG-based || Grammar || Unification92.889.5889.64
Model || SHRG-based || Grammar || Rewriting-90.3989.51
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Parsing accuracy of the best existing grammar-free and -based models as well as our SHRG-based model. Results are copied from (Oepen et al., 2015; Peng et al., 2017a; Buys and Blunsom, 2017).}\n\\begin{tabular}{llrr}\n\\toprule\n & SDG & EDS & DMRS \\\\\n\\midrule\nModel $\\parallel$ Data-driven $\\parallel$ Grammar $\\parallel$ NO & 89.4 & 85.48 & 84.16 \\\\\nModel $\\parallel$ ERG-based $\\parallel$ Grammar $\\parallel$ Unification & 92.8 & 89.58 & 89.64 \\\\\nModel $\\parallel$ SHRG-based $\\parallel$ Grammar $\\parallel$ Rewriting & - & 90.39 & 89.51 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Parsing accuracy of the best existing grammar-free and -based models as well as our SHRG-based model. Results are copied from (Oepen et al., 2015; Peng et al., 2017a; Buys and Blunsom, 2017).\n
\n \n SDG\n \n EDS\n \n DMRS\n
\n Model || Data-driven || Grammar || NO\n \n 89.4\n \n 85.48\n \n 84.16\n
\n Model || ERG-based || Grammar || Unification\n \n 92.8\n \n 89.58\n \n 89.64\n
\n Model || SHRG-based || Grammar || Rewriting\n \n -\n \n 90.39\n \n 89.51\n
\n" }, { "instance_id": "01449897-6066-4d15-8c10-5f49f14fef78", "table_id_paper": "P18-1048table_2", "caption": "Detection performance (trigger identification plus multi-class classification)", "row_header_level": 2, "row_headers": [ [ "Methods", "MSEP-EMD" ], [ "Methods", "Cross-Event" ], [ "Methods", "Cross-Entity" ], [ "Methods", "Joint (Local+Global)" ], [ "Methods", "CNN" ], [ "Methods", "DM-CNN" ], [ "Methods", "NC-CNN" ], [ "Methods", "FB-RNN (GRU)" ], [ "Methods", "Bi-RNN (GRU)" ], [ "Methods", "ANNs (ACE+FN)" ], [ "Methods", "DM-CNN?(ACE+Wiki)" ], [ "Methods", "ANN-S2 (ACE+FN)" ], [ "Methods", "Hybrid: Bi-LSTM+CNN" ], [ "Methods", "SELF: Bi-LSTM+GAN" ] ], "column_header_level": 1, "column_headers": [ [ "P (%)" ], [ "R (%)" ], [ "F (%)" ] ], "contents": [ [ "70.4", "65", "67.6" ], [ "68.8", "68.9", "68.8" ], [ "72.9", "64.3", "68.3" ], [ "73.7", "62.3", "67.5" ], [ "71.8", "66.4", "69" ], [ "75.6", "63.6", "69.1" ], [ "-", "-", "71.3" ], [ "66.8", "68", "67.4" ], [ "66", "73", "69.3" ], [ "77.6", "65.2", "70.7" ], [ "75.7", "66", "70.5" ], [ "76.8", "67.5", "71.9" ], [ "84.6", "64.9", "73.4" ], [ "71.3", "74.7", "73" ] ], "metrics_loc": "column", "metrics_type": "['P (%)', 'R (%)', 'F (%)']", "target_entity": "['SELF: Bi-LSTM+GAN']", "table_html": "
P (%) R (%) F (%)
Methods || MSEP-EMD 70.4 65 67.6
Methods || Cross-Event 68.8 68.9 68.8
Methods || Cross-Entity 72.9 64.3 68.3
Methods || Joint (Local+Global) 73.7 62.3 67.5
Methods || CNN 71.8 66.4 69
Methods || DM-CNN 75.6 63.6 69.1
Methods || NC-CNN - - 71.3
Methods || FB-RNN (GRU) 66.8 68 67.4
Methods || Bi-RNN (GRU) 66 73 69.3
Methods || ANNs (ACE+FN) 77.6 65.2 70.7
Methods || DM-CNN?(ACE+Wiki) 75.7 66 70.5
Methods || ANN-S2 (ACE+FN) 76.8 67.5 71.9
Methods || Hybrid: Bi-LSTM+CNN 84.6 64.9 73.4
Methods || SELF: Bi-LSTM+GAN 71.3 74.7 73
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P18-1048", "page_no": 6, "dir": "acl2018", "description": "Table 2 shows the performance of multi-class classification. SELF achieves nearly the same F-score as Feng et al (2016)'s Hybrid, and outperforms the others. More importantly, SELF is the only one which obtains a performance higher than 70% for both precision and recall. Besides, by analyzing the experimental results, we have identified the following regularities:. ? Similar to the pattern classifiers that are based on hand-designed features, the CNN models enable higher precision to be obtained. However the recall is lower. ? The RNN models contribute to achieving a higher recall. However the precision is lower. ? Expansion of the training data set helps to increase the precision. Let us turn to the structurally more complicated models, SELF and Hybrid. SELF inherits the merits of the RNN models, classifying the events with higher recall. Besides, by the utilization of GAN, SELF has evolved from the traditional learning strategies, being capable of learning from GAN and getting rid of the mistakenly generated spurious features. So that it outperforms other RNNs, with improvements of no less than 4.5% precision and 1.7% recall.", "class_sentence": "[1, 1, 1, 2, 1, 1, 1, 1, 2, 1, 1, 2, 1]", "sentences": "['Table 2 shows the performance of multi-class classification.', \"SELF achieves nearly the same F-score as Feng et al (2016)'s Hybrid, and outperforms the others.\", 'More importantly, SELF is the only one which obtains a performance higher than 70% for both precision and recall.', 'Besides, by analyzing the experimental results, we have identified the following regularities:.', '? Similar to the pattern classifiers that are based on hand-designed features, the CNN models enable higher precision to be obtained.', 'However the recall is lower.', '? The RNN models contribute to achieving a higher recall.', 'However the precision is lower.', '? Expansion of the training data set helps to increase the precision.', 'Let us turn to the structurally more complicated models, SELF and Hybrid.', 'SELF inherits the merits of the RNN models, classifying the events with higher recall.', 'Besides, by the utilization of GAN, SELF has evolved from the traditional learning strategies, being capable of learning from GAN and getting rid of the mistakenly generated spurious features.', 'So that it outperforms other RNNs, with improvements of no less than 4.5% precision and 1.7% recall.']", "header_mention": "[None, ['SELF: Bi-LSTM+GAN', 'Hybrid: Bi-LSTM+CNN', 'P (%)', 'R (%)', 'F (%)'], ['SELF: Bi-LSTM+GAN', 'P (%)', 'R (%)'], None, ['CNN', 'DM-CNN', 'P (%)'], ['CNN', 'DM-CNN', 'R (%)'], ['FB-RNN (GRU)', 'Bi-RNN (GRU)', 'R (%)'], ['FB-RNN (GRU)', 'Bi-RNN (GRU)', 'P (%)'], ['P (%)'], ['Hybrid: Bi-LSTM+CNN', 'SELF: Bi-LSTM+GAN'], ['SELF: Bi-LSTM+GAN', 'FB-RNN (GRU)', 'Bi-RNN (GRU)', 'R (%)'], ['SELF: Bi-LSTM+GAN'], ['FB-RNN (GRU)', 'SELF: Bi-LSTM+GAN', 'R (%)']]", "valid": 1, "image_id": "P18-1048-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Detection performance (trigger identification plus multi-class classification)

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n P (%)R (%)F (%)
Methods || MSEP-EMD70.46567.6
Methods || Cross-Event68.868.968.8
Methods || Cross-Entity72.964.368.3
Methods || Joint (Local+Global)73.762.367.5
Methods || CNN71.866.469
Methods || DM-CNN75.663.669.1
Methods || NC-CNN--71.3
Methods || FB-RNN (GRU)66.86867.4
Methods || Bi-RNN (GRU)667369.3
Methods || ANNs (ACE+FN)77.665.270.7
Methods || DM-CNN?(ACE+Wiki)75.76670.5
Methods || ANN-S2 (ACE+FN)76.867.571.9
Methods || Hybrid: Bi-LSTM+CNN84.664.973.4
Methods || SELF: Bi-LSTM+GAN71.374.773
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Detection performance (trigger identification plus multi-class classification)}\n\\begin{tabular}{lllr}\n\\toprule\n & P (\\%) & R (\\%) & F (\\%) \\\\\n\\midrule\nMethods $\\parallel$ MSEP-EMD & 70.4 & 65 & 67.6 \\\\\nMethods $\\parallel$ Cross-Event & 68.8 & 68.9 & 68.8 \\\\\nMethods $\\parallel$ Cross-Entity & 72.9 & 64.3 & 68.3 \\\\\nMethods $\\parallel$ Joint (Local+Global) & 73.7 & 62.3 & 67.5 \\\\\nMethods $\\parallel$ CNN & 71.8 & 66.4 & 69 \\\\\nMethods $\\parallel$ DM-CNN & 75.6 & 63.6 & 69.1 \\\\\nMethods $\\parallel$ NC-CNN & - & - & 71.3 \\\\\nMethods $\\parallel$ FB-RNN (GRU) & 66.8 & 68 & 67.4 \\\\\nMethods $\\parallel$ Bi-RNN (GRU) & 66 & 73 & 69.3 \\\\\nMethods $\\parallel$ ANNs (ACE+FN) & 77.6 & 65.2 & 70.7 \\\\\nMethods $\\parallel$ DM-CNN?(ACE+Wiki) & 75.7 & 66 & 70.5 \\\\\nMethods $\\parallel$ ANN-S2 (ACE+FN) & 76.8 & 67.5 & 71.9 \\\\\nMethods $\\parallel$ Hybrid: Bi-LSTM+CNN & 84.6 & 64.9 & 73.4 \\\\\nMethods $\\parallel$ SELF: Bi-LSTM+GAN & 71.3 & 74.7 & 73 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Detection performance (trigger identification plus multi-class classification)\n
\n \n P (%)\n \n R (%)\n \n F (%)\n
\n Methods || MSEP-EMD\n \n 70.4\n \n 65\n \n 67.6\n
\n Methods || Cross-Event\n \n 68.8\n \n 68.9\n \n 68.8\n
\n Methods || Cross-Entity\n \n 72.9\n \n 64.3\n \n 68.3\n
\n Methods || Joint (Local+Global)\n \n 73.7\n \n 62.3\n \n 67.5\n
\n Methods || CNN\n \n 71.8\n \n 66.4\n \n 69\n
\n Methods || DM-CNN\n \n 75.6\n \n 63.6\n \n 69.1\n
\n Methods || NC-CNN\n \n -\n \n -\n \n 71.3\n
\n Methods || FB-RNN (GRU)\n \n 66.8\n \n 68\n \n 67.4\n
\n Methods || Bi-RNN (GRU)\n \n 66\n \n 73\n \n 69.3\n
\n Methods || ANNs (ACE+FN)\n \n 77.6\n \n 65.2\n \n 70.7\n
\n Methods || DM-CNN?(ACE+Wiki)\n \n 75.7\n \n 66\n \n 70.5\n
\n Methods || ANN-S2 (ACE+FN)\n \n 76.8\n \n 67.5\n \n 71.9\n
\n Methods || Hybrid: Bi-LSTM+CNN\n \n 84.6\n \n 64.9\n \n 73.4\n
\n Methods || SELF: Bi-LSTM+GAN\n \n 71.3\n \n 74.7\n \n 73\n
\n" }, { "instance_id": "eee63a8f-bf0c-4c17-ba99-5b15b8182e46", "table_id_paper": "P18-1050table_6", "caption": "Results on MCNC task", "row_header_level": 2, "row_headers": [ [ "Models", "(Chambers and Jurafsky 2008)" ], [ "Models", "(Granroth-Wilding and Clark 2016)" ], [ "Models", "(Pichotta and Mooney 2016)" ], [ "Models", "(Wang et al. 2017)" ], [ "Models", "Our Results" ] ], "column_header_level": 1, "column_headers": [ [ "Acc.(%)" ] ], "contents": [ [ "30.92" ], [ "43.28" ], [ "43.17" ], [ "46.67" ], [ "48.83" ] ], "metrics_loc": "column", "metrics_type": "['Acc.(%)']", "target_entity": "['Our Results']", "table_html": "
Acc.(%)
Models || (Chambers and Jurafsky 2008) 30.92
Models || (Granroth-Wilding and Clark 2016) 43.28
Models || (Pichotta and Mooney 2016) 43.17
Models || (Wang et al. 2017) 46.67
Models || Our Results 48.83
", "table_name": "Table 6", "table_id": "table_6", "paper_id": "P18-1050", "page_no": 8, "dir": "acl2018", "description": "Table 6 shows the comparisons of our results with the performance of several previous models, which were all trained with 1,500k event chains extracted from the NYT portion of the Gigaword corpus (Graff and Cieri 2003). Each event chain consists of a sequence of verbs sharing an actor within a news article. Except Chambers and Jurafsky (2008), other recent models utilized more and more sophisticated neural language models. Granroth-Wilding and Clark (2016) proposed a two layer neural network model that learns embeddings of event predicates and their arguments for predicting the next event. Pichotta and Mooney (2016) introduced a LSTM-based language model for event prediction. Wang et al.(2017) used dynamic memory as attention in LSTM for prediction. It is encouraging that by using event knowledge extracted from automatically identified narratives, we achieved the best event prediction performance, which is 2.2% higher than the best neural network model.", "class_sentence": "[1, 2, 2, 2, 2, 2, 1]", "sentences": "['Table 6 shows the comparisons of our results with the performance of several previous models, which were all trained with 1,500k event chains extracted from the NYT portion of the Gigaword corpus (Graff and Cieri 2003).', 'Each event chain consists of a sequence of verbs sharing an actor within a news article.', 'Except Chambers and Jurafsky (2008), other recent models utilized more and more sophisticated neural language models.', 'Granroth-Wilding and Clark (2016) proposed a two layer neural network model that learns embeddings of event predicates and their arguments for predicting the next event.', 'Pichotta and Mooney (2016) introduced a LSTM-based language model for event prediction.', 'Wang et al.(2017) used dynamic memory as attention in LSTM for prediction.', 'It is encouraging that by using event knowledge extracted from automatically identified narratives, we achieved the best event prediction performance, which is 2.2% higher than the best neural network model.']", "header_mention": "[None, None, ['(Granroth-Wilding and Clark 2016)', '(Pichotta and Mooney 2016)', '(Wang et al. 2017)'], ['(Granroth-Wilding and Clark 2016)'], ['(Pichotta and Mooney 2016)'], ['(Wang et al. 2017)'], ['(Wang et al. 2017)', 'Our Results']]", "valid": 1, "image_id": "P18-1050-Table6-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results on MCNC task

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Acc.(%)
Models || (Chambers and Jurafsky 2008)30.92
Models || (Granroth-Wilding and Clark 2016)43.28
Models || (Pichotta and Mooney 2016)43.17
Models || (Wang et al. 2017)46.67
Models || Our Results48.83
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{5}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results on MCNC task}\n\\begin{tabular}{lr}\n\\toprule\n & Acc.(\\%) \\\\\n\\midrule\nModels $\\parallel$ (Chambers and Jurafsky 2008) & 30.92 \\\\\nModels $\\parallel$ (Granroth-Wilding and Clark 2016) & 43.28 \\\\\nModels $\\parallel$ (Pichotta and Mooney 2016) & 43.17 \\\\\nModels $\\parallel$ (Wang et al. 2017) & 46.67 \\\\\nModels $\\parallel$ Our Results & 48.83 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 6: Results on MCNC task\n
\n \n Acc.(%)\n
\n Models || (Chambers and Jurafsky 2008)\n \n 30.92\n
\n Models || (Granroth-Wilding and Clark 2016)\n \n 43.28\n
\n Models || (Pichotta and Mooney 2016)\n \n 43.17\n
\n Models || (Wang et al. 2017)\n \n 46.67\n
\n Models || Our Results\n \n 48.83\n
\n" }, { "instance_id": "d74edf18-a763-40b8-b296-134f136c5ee6", "table_id_paper": "P18-1076table_5", "caption": "Results for key-value knowledge retrieval (CN5Sel, 50 facts). Subj/Obj and integration. means: we attend over the fact subject (Key) and take the weighted fact object as value (Value). CN", "row_header_level": 2, "row_headers": [ [ "Key/Value", "Subj/Obj" ], [ "Key/Value", "Obj/Obj" ] ], "column_header_level": 2, "column_headers": [ [ "NE", "Dev" ], [ "NE", "Test" ], [ "CN", "Dev" ], [ "CN", "Test" ] ], "contents": [ [ "76.65", "71.52", "71.85", "67.64" ], [ "76.7", "71.28", "71.25", "67.48" ] ], "metrics_loc": "column", "metrics_type": "['accurary', 'accurary', 'accurary', 'accurary']", "target_entity": "['Subj/Obj']", "table_html": "
NE || Dev NE || Test CN || Dev CN || Test
Key/Value || Subj/Obj 76.65 71.52 71.85 67.64
Key/Value || Obj/Obj 76.7 71.28 71.25 67.48
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "P18-1076", "page_no": 7, "dir": "acl2018", "description": "Table 5 shows that for the NE dataset, the two strategies performequally well on the Dev set, whereas the Subj/Obj strategy works slightly better on the Test set. For Common Nouns, Subj/Obj is better.", "class_sentence": "[1, 1]", "sentences": "['Table 5 shows that for the NE dataset, the two strategies performequally well on the Dev set, whereas the Subj/Obj strategy works slightly better on the Test set.', 'For Common Nouns, Subj/Obj is better.']", "header_mention": "[['NE', 'Subj/Obj', 'Obj/Obj', 'Dev', 'Test'], ['CN', 'Subj/Obj']]", "valid": 1, "image_id": "P18-1076-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results for key-value knowledge retrieval (CN5Sel, 50 facts). Subj/Obj and integration. means: we attend over the fact subject (Key) and take the weighted fact object as value (Value). CN

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n NE || DevNE || TestCN || DevCN || Test
Key/Value || Subj/Obj76.6571.5271.8567.64
Key/Value || Obj/Obj76.771.2871.2567.48
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results for key-value knowledge retrieval (CN5Sel, 50 facts). Subj/Obj and integration. means: we attend over the fact subject (Key) and take the weighted fact object as value (Value). CN}\n\\begin{tabular}{lrrrr}\n\\toprule\n & NE $\\parallel$ Dev & NE $\\parallel$ Test & CN $\\parallel$ Dev & CN $\\parallel$ Test \\\\\n\\midrule\nKey/Value $\\parallel$ Subj/Obj & 76.65 & 71.52 & 71.85 & 67.64 \\\\\nKey/Value $\\parallel$ Obj/Obj & 76.7 & 71.28 & 71.25 & 67.48 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Results for key-value knowledge retrieval (CN5Sel, 50 facts). Subj/Obj and integration. means: we attend over the fact subject (Key) and take the weighted fact object as value (Value). CN\n
\n \n NE || Dev\n \n NE || Test\n \n CN || Dev\n \n CN || Test\n
\n Key/Value || Subj/Obj\n \n 76.65\n \n 71.52\n \n 71.85\n \n 67.64\n
\n Key/Value || Obj/Obj\n \n 76.7\n \n 71.28\n \n 71.25\n \n 67.48\n
\n" }, { "instance_id": "03edbc45-b8d5-429e-8919-bc6e0d499a98", "table_id_paper": "P18-1101table_4", "caption": "Human evaluation results on judging the homogeneity of latent actions in SMD.", "row_header_level": 2, "row_headers": [ [ "Model", "DI-VAE" ], [ "Model", "DI-VST" ] ], "column_header_level": 1, "column_headers": [ [ "Exp Agree" ], [ "Worker \u03ba" ], [ "Match Rate" ] ], "contents": [ [ "85.60%", "0.52", "71.30%" ], [ "93.30%", "0.48", "74.90%" ] ], "metrics_loc": "column", "metrics_type": "['Exp Agree', 'Worker \u03ba', 'Match Rate']", "target_entity": "['DI-VST']", "table_html": "
Exp Agree Worker \u03ba Match Rate
Model || DI-VAE 85.60% 0.52 71.30%
Model || DI-VST 93.30% 0.48 74.90%
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P18-1101", "page_no": 7, "dir": "acl2018", "description": "Table 4 shows that both methods work well and DI-VST achieved better homogeneity than DI-VAE.", "class_sentence": "[1]", "sentences": "['Table 4 shows that both methods work well and DI-VST achieved better homogeneity than DI-VAE.']", "header_mention": "[['DI-VST', 'DI-VAE']]", "valid": 1, "image_id": "P18-1101-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Human evaluation results on judging the homogeneity of latent actions in SMD.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Exp AgreeWorker \u03baMatch Rate
Model || DI-VAE85.60%0.5271.30%
Model || DI-VST93.30%0.4874.90%
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Human evaluation results on judging the homogeneity of latent actions in SMD.}\n\\begin{tabular}{llrl}\n\\toprule\n & Exp Agree & Worker $\\kappa$ & Match Rate \\\\\n\\midrule\nModel $\\parallel$ DI-VAE & 85.60\\% & 0.52 & 71.30\\% \\\\\nModel $\\parallel$ DI-VST & 93.30\\% & 0.48 & 74.90\\% \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Human evaluation results on judging the homogeneity of latent actions in SMD.\n
\n \n Exp Agree\n \n Worker \u03ba\n \n Match Rate\n
\n Model || DI-VAE\n \n 85.60%\n \n 0.52\n \n 71.30%\n
\n Model || DI-VST\n \n 93.30%\n \n 0.48\n \n 74.90%\n
\n" }, { "instance_id": "5e9034ef-6ea5-40bd-aa85-b5fe5620f7e3", "table_id_paper": "P18-1102table_4", "caption": "Results on the human evaluation.", "row_header_level": 1, "row_headers": [ [ "Seq2Seq-att" ], [ "MMI-bidi" ], [ "MARM" ], [ "Seq2Seq+IDF" ], [ "SC-Seq2SeqNIWF s=1" ], [ "SC-Seq2SeqNIWF s=0.5" ], [ "SC-Seq2SeqNIWF s=0" ] ], "column_header_level": 1, "column_headers": [ [ "+2" ], [ "+1" ], [ "+0" ], [ "kappa" ] ], "contents": [ [ "29.32%", "25.27%", "45.41%", "0.448" ], [ "30.40%", "24.85%", "44.75%", "0.471" ], [ "20.11%", "27.96%", "51.93%", "0.404" ], [ "28.81%", "23.87%", "47.33%", "0.418" ], [ "42.47%", "14.29%", "43.24%", "0.507" ], [ "20.62%", "40.16%", "39.22%", "0.451" ], [ "14.34%", "46.38%", "39.28%", "0.526" ] ], "metrics_loc": "column", "metrics_type": "['+2', '+1', '+0', 'kappa']", "target_entity": "['SC-Seq2SeqNIWF s=1', 'SC-Seq2SeqNIWF s=0.5', 'SC-Seq2SeqNIWF s=0']", "table_html": "
+2 +1 +0 kappa
Seq2Seq-att 29.32% 25.27% 45.41% 0.448
MMI-bidi 30.40% 24.85% 44.75% 0.471
MARM 20.11% 27.96% 51.93% 0.404
Seq2Seq+IDF 28.81% 23.87% 47.33% 0.418
SC-Seq2SeqNIWF s=1 42.47% 14.29% 43.24% 0.507
SC-Seq2SeqNIWF s=0.5 20.62% 40.16% 39.22% 0.451
SC-Seq2SeqNIWF s=0 14.34% 46.38% 39.28% 0.526
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P18-1102", "page_no": 8, "dir": "acl2018", "description": "Table 4 shows the human evaluation results. We can observe that: (1) SC-Seq2SeqNIWF s=1 generates the most informative responses and interesting (labeled as \u0081g+2\u0081h) and the least general responses than all the baseline models. Meanwhile, SC-Seq2SeqNIWF s=0 generates the most general responses (labeled as \u0081g+1\u0081h); (2) MARM generates the most bad responses (labeled as \u0081g+0\u0081h), which indicates the drawbacks of the unknown latent responding mechanisms; (3) The kappa values of our models are all larger than 0.4, considered as \u0081gmoderate agreement\u0081h regarding quality of responses. The largest kappa value is achieved by SC-Seq2SeqNIWF s=0, which seems reasonable since it is easy to reach an agreement on general responses. Sign tests demonstrate the improvements of SC-Seq2SeqNIWF s=1 to the baseline models are statistically significant (p-value < 0.01). All the human judgement results again demonstrate the effectiveness of our controlled generation mechanism.", "class_sentence": "[1, 1, 1, 1, 2, 2]", "sentences": "['Table 4 shows the human evaluation results.', 'We can observe that: (1) SC-Seq2SeqNIWF s=1 generates the most informative responses and interesting (labeled as \\x81g+2\\x81h) and the least general responses than all the baseline models.', 'Meanwhile, SC-Seq2SeqNIWF s=0 generates the most general responses (labeled as \\x81g+1\\x81h); (2) MARM generates the most bad responses (labeled as \\x81g+0\\x81h), which indicates the drawbacks of the unknown latent responding mechanisms; (3) The kappa values of our models are all larger than 0.4, considered as \\x81gmoderate agreement\\x81h regarding quality of responses.', 'The largest kappa value is achieved by SC-Seq2SeqNIWF s=0, which seems reasonable since it is easy to reach an agreement on general responses.', 'Sign tests demonstrate the improvements of SC-Seq2SeqNIWF s=1 to the baseline models are statistically significant (p-value < 0.01).', 'All the human judgement results again demonstrate the effectiveness of our controlled generation mechanism.']", "header_mention": "[None, ['SC-Seq2SeqNIWF s=1', '+2'], ['SC-Seq2SeqNIWF s=0', '+1', 'kappa', 'SC-Seq2SeqNIWF s=1', 'SC-Seq2SeqNIWF s=0.5'], ['kappa', 'SC-Seq2SeqNIWF s=0'], None, None]", "valid": 1, "image_id": "P18-1102-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results on the human evaluation.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n +2+1+0kappa
Seq2Seq-att29.32%25.27%45.41%0.448
MMI-bidi30.40%24.85%44.75%0.471
MARM20.11%27.96%51.93%0.404
Seq2Seq+IDF28.81%23.87%47.33%0.418
SC-Seq2SeqNIWF s=142.47%14.29%43.24%0.507
SC-Seq2SeqNIWF s=0.520.62%40.16%39.22%0.451
SC-Seq2SeqNIWF s=014.34%46.38%39.28%0.526
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results on the human evaluation.}\n\\begin{tabular}{llllr}\n\\toprule\n & +2 & +1 & +0 & kappa \\\\\n\\midrule\nSeq2Seq-att & 29.32\\% & 25.27\\% & 45.41\\% & 0.448 \\\\\nMMI-bidi & 30.40\\% & 24.85\\% & 44.75\\% & 0.471 \\\\\nMARM & 20.11\\% & 27.96\\% & 51.93\\% & 0.404 \\\\\nSeq2Seq+IDF & 28.81\\% & 23.87\\% & 47.33\\% & 0.418 \\\\\nSC-Seq2SeqNIWF s$=$1 & 42.47\\% & 14.29\\% & 43.24\\% & 0.507 \\\\\nSC-Seq2SeqNIWF s$=$0.5 & 20.62\\% & 40.16\\% & 39.22\\% & 0.451 \\\\\nSC-Seq2SeqNIWF s$=$0 & 14.34\\% & 46.38\\% & 39.28\\% & 0.526 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Results on the human evaluation.\n
\n \n +2\n \n +1\n \n +0\n \n kappa\n
\n Seq2Seq-att\n \n 29.32%\n \n 25.27%\n \n 45.41%\n \n 0.448\n
\n MMI-bidi\n \n 30.40%\n \n 24.85%\n \n 44.75%\n \n 0.471\n
\n MARM\n \n 20.11%\n \n 27.96%\n \n 51.93%\n \n 0.404\n
\n Seq2Seq+IDF\n \n 28.81%\n \n 23.87%\n \n 47.33%\n \n 0.418\n
\n SC-Seq2SeqNIWF s=1\n \n 42.47%\n \n 14.29%\n \n 43.24%\n \n 0.507\n
\n SC-Seq2SeqNIWF s=0.5\n \n 20.62%\n \n 40.16%\n \n 39.22%\n \n 0.451\n
\n SC-Seq2SeqNIWF s=0\n \n 14.34%\n \n 46.38%\n \n 39.28%\n \n 0.526\n
\n" }, { "instance_id": "7fd03b33-0082-400e-bae8-3219a56253ce", "table_id_paper": "P18-1145table_2", "caption": "Experiment results on ACE2005 and KBPEval2017. * indicates the result adapted from the original paper. For KBPEval2017, \u201cTrigger Identification\u201d corresponds to the \u201cSpan\u201d metric and \u201cTrigger Classification\u201d corresponds to the \u201cType\u201d metric reported in official evaluation.", "row_header_level": 2, "row_headers": [ [ "Model", "FBRNN(Char)" ], [ "Model", "DMCNN(Char)" ], [ "Model", "C-BiLSTM*" ], [ "Model", "FBRNN(Word)" ], [ "Model", "DMCNN(Word)" ], [ "Model", "HNN*" ], [ "Model", "Rich-C*" ], [ "Model", "KBP2017 Best*" ], [ "Model", "NPN(Concat)" ], [ "Model", "NPN(General)" ], [ "Model", "NPN(Task-specific)" ] ], "column_header_level": 3, "column_headers": [ [ "ACE2005", "Trigger Identification", "P" ], [ "ACE2005", "Trigger Identification", "R" ], [ "ACE2005", "Trigger Identification", "F1" ], [ "ACE2005", "Trigger Classification", "P" ], [ "ACE2005", "Trigger Classification", "R" ], [ "ACE2005", "Trigger Classification", "F1" ], [ "KBPEval2017", "Trigger Identification", "P" ], [ "KBPEval2017", "Trigger Identification", "R" ], [ "KBPEval2017", "Trigger Identification", "F1" ], [ "KBPEval2017", "Trigger Classification", "P" ], [ "KBPEval2017", "Trigger Classification", "R" ], [ "KBPEval2017", "Trigger Classification", "F1" ] ], "contents": [ [ "61.3", "45.6", "52.3", "57.5", "42.8", "49.1", "57.97", "36.92", "45.11", "51.71", "32.94", "40.24" ], [ "60.1", "61.6", "60.9", "57.1", "58.5", "57.8", "53.67", "49.92", "51.73", "50.03", "46.53", "48.22" ], [ "65.6", "66.7", "66.1", "60.0", "60.9", "60.4", "-", "-", "-", "-", "-", "-" ], [ "64.1", "63.7", "63.9", "59.9", "59.6", "59.7", "65.10", "46.86", "54.50", "60.05", "43.22", "50.27" ], [ "66.6", "63.6", "65.1", "61.6", "58.8", "60.2", "60.43", "51.64", "55.69", "54.81", "46.84", "50.51" ], [ "74.2", "63.1", "68.2", "77.1", "53.1", "63.0", "-", "-", "-", "-", "-", "-" ], [ "62.2", "71.9", "66.7", "58.9", "68.1", "63.2", "-", "-", "-", "-", "-", "-" ], [ "-", "-", "-", "-", "-", "-", "67.76", "45.92", "54.74", "62.69", "42.48", "50.64" ], [ "76.5", "59.8", "67.1", "72.8", "56.9", "63.9", "64.58", "50.31", "56.56", "59.14", "46.07", "51.80" ], [ "71.5", "63.2", "67.1", "67.3", "59.6", "63.2", "63.67", "51.32", "56.83", "57.78", "46.58", "51.57" ], [ "64.8", "73.8", "69.0", "60.9", "69.3", "64.8", "64.32", "53.16", "58.21", "57.63", "47.63", "52.15" ] ], "metrics_loc": "column", "metrics_type": "['P', 'R', 'F1', 'P', 'R', 'F1', 'P', 'R', 'F1', 'P', 'R', 'F1']", "target_entity": "['NPN(Task-specific)']", "table_html": "
ACE2005 || Trigger Identification || P ACE2005 || Trigger Identification || R ACE2005 || Trigger Identification || F1 ACE2005 || Trigger Classification || P ACE2005 || Trigger Classification || R ACE2005 || Trigger Classification || F1 KBPEval2017 || Trigger Identification || P KBPEval2017 || Trigger Identification || R KBPEval2017 || Trigger Identification || F1 KBPEval2017 || Trigger Classification || P KBPEval2017 || Trigger Classification || R KBPEval2017 || Trigger Classification || F1
Model || FBRNN(Char) 61.3 45.6 52.3 57.5 42.8 49.1 57.97 36.92 45.11 51.71 32.94 40.24
Model || DMCNN(Char) 60.1 61.6 60.9 57.1 58.5 57.8 53.67 49.92 51.73 50.03 46.53 48.22
Model || C-BiLSTM* 65.6 66.7 66.1 60.0 60.9 60.4 - - - - - -
Model || FBRNN(Word) 64.1 63.7 63.9 59.9 59.6 59.7 65.10 46.86 54.50 60.05 43.22 50.27
Model || DMCNN(Word) 66.6 63.6 65.1 61.6 58.8 60.2 60.43 51.64 55.69 54.81 46.84 50.51
Model || HNN* 74.2 63.1 68.2 77.1 53.1 63.0 - - - - - -
Model || Rich-C* 62.2 71.9 66.7 58.9 68.1 63.2 - - - - - -
Model || KBP2017 Best* - - - - - - 67.76 45.92 54.74 62.69 42.48 50.64
Model || NPN(Concat) 76.5 59.8 67.1 72.8 56.9 63.9 64.58 50.31 56.56 59.14 46.07 51.80
Model || NPN(General) 71.5 63.2 67.1 67.3 59.6 63.2 63.67 51.32 56.83 57.78 46.58 51.57
Model || NPN(Task-specific) 64.8 73.8 69.0 60.9 69.3 64.8 64.32 53.16 58.21 57.63 47.63 52.15
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P18-1145", "page_no": 6, "dir": "acl2018", "description": "Table 2 shows the results on ACE2005 and KBPEval2017. From this table, we can see that:. 1) NPNs steadily outperform all baselines significantly. Compared with baselines, NPN(Taskspecific) gains at least 1.6 (2.5%) and 1.5 (3.0%) F1-score improvements on trigger classification task on ACE2005 and KBPEval2017 respectively. 2) By exploiting compositional structures of triggers, our trigger nugget generator can effectively resolve the word-trigger mismatch problem. As shown in Table 2, NPN(Taskspecific) achieved significant F1-score improvements on trigger identification task on both datasets. It is notable that our method achieved a remarkable high recall on both datasets, which indicates that NPNs do detect a number of triggers which previous methods can not identify. 3) By summarizing information from both characters and words, the hybrid representation learning is effective for event detection. Comparing with corresponding characterbased methods3 , word-based methods achieved 2 to 3 F1-score improvements, which indicates that words can provide additional information for event detection. By combining character-level and word-level features, NPNs are able to perform character-based event detection meanwhile take word-level knowledge into consideration too.", "class_sentence": "[1, 1, 1, 1, 2, 1, 2, 2, 1, 2]", "sentences": "['Table 2 shows the results on ACE2005 and KBPEval2017.', 'From this table, we can see that:.', '1) NPNs steadily outperform all baselines significantly.', 'Compared with baselines, NPN(Taskspecific) gains at least 1.6 (2.5%) and 1.5 (3.0%) F1-score improvements on trigger classification task on ACE2005 and KBPEval2017 respectively.', '2) By exploiting compositional structures of triggers, our trigger nugget generator can effectively resolve the word-trigger mismatch problem.', 'As shown in Table 2, NPN(Taskspecific) achieved significant F1-score improvements on trigger identification task on both datasets.', 'It is notable that our method achieved a remarkable high recall on both datasets, which indicates that NPNs do detect a number of triggers which previous methods can not identify.', '3) By summarizing information from both characters and words, the hybrid representation learning is effective for event detection.', 'Comparing with corresponding characterbased methods3 , word-based methods achieved 2 to 3 F1-score improvements, which indicates that words can provide additional information for event detection.', 'By combining character-level and word-level features, NPNs are able to perform character-based event detection meanwhile take word-level knowledge into consideration too.']", "header_mention": "[['ACE2005', 'KBPEval2017'], None, ['NPN(Concat)', 'NPN(General)', 'NPN(Task-specific)'], ['NPN(Task-specific)', 'ACE2005', 'KBPEval2017', 'F1'], ['Trigger Identification', 'Trigger Classification'], ['NPN(Task-specific)', 'Trigger Identification', 'Trigger Classification'], ['NPN(Task-specific)'], None, ['NPN(Concat)', 'NPN(General)', 'NPN(Task-specific)', 'C-BiLSTM*', 'FBRNN(Char)', 'DMCNN(Char)', 'F1'], ['NPN(Concat)', 'NPN(General)', 'NPN(Task-specific)']]", "valid": 1, "image_id": "P18-1145-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Experiment results on ACE2005 and KBPEval2017. * indicates the result adapted from the original paper. For KBPEval2017, \u201cTrigger Identification\u201d corresponds to the \u201cSpan\u201d metric and \u201cTrigger Classification\u201d corresponds to the \u201cType\u201d metric reported in official evaluation.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n ACE2005 || Trigger Identification || PACE2005 || Trigger Identification || RACE2005 || Trigger Identification || F1ACE2005 || Trigger Classification || PACE2005 || Trigger Classification || RACE2005 || Trigger Classification || F1KBPEval2017 || Trigger Identification || PKBPEval2017 || Trigger Identification || RKBPEval2017 || Trigger Identification || F1KBPEval2017 || Trigger Classification || PKBPEval2017 || Trigger Classification || RKBPEval2017 || Trigger Classification || F1
Model || FBRNN(Char)61.345.652.357.542.849.157.9736.9245.1151.7132.9440.24
Model || DMCNN(Char)60.161.660.957.158.557.853.6749.9251.7350.0346.5348.22
Model || C-BiLSTM*65.666.766.160.060.960.4------
Model || FBRNN(Word)64.163.763.959.959.659.765.1046.8654.5060.0543.2250.27
Model || DMCNN(Word)66.663.665.161.658.860.260.4351.6455.6954.8146.8450.51
Model || HNN*74.263.168.277.153.163.0------
Model || Rich-C*62.271.966.758.968.163.2------
Model || KBP2017 Best*------67.7645.9254.7462.6942.4850.64
Model || NPN(Concat)76.559.867.172.856.963.964.5850.3156.5659.1446.0751.80
Model || NPN(General)71.563.267.167.359.663.263.6751.3256.8357.7846.5851.57
Model || NPN(Task-specific)64.873.869.060.969.364.864.3253.1658.2157.6347.6352.15
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Experiment results on ACE2005 and KBPEval2017. * indicates the result adapted from the original paper. For KBPEval2017, \u201cTrigger Identification\u201d corresponds to the \u201cSpan\u201d metric and \u201cTrigger Classification\u201d corresponds to the \u201cType\u201d metric reported in official evaluation.}\n\\begin{tabular}{lllllllllllll}\n\\toprule\n & ACE2005 $\\parallel$ Trigger Identification $\\parallel$ P & ACE2005 $\\parallel$ Trigger Identification $\\parallel$ R & ACE2005 $\\parallel$ Trigger Identification $\\parallel$ F1 & ACE2005 $\\parallel$ Trigger Classification $\\parallel$ P & ACE2005 $\\parallel$ Trigger Classification $\\parallel$ R & ACE2005 $\\parallel$ Trigger Classification $\\parallel$ F1 & KBPEval2017 $\\parallel$ Trigger Identification $\\parallel$ P & KBPEval2017 $\\parallel$ Trigger Identification $\\parallel$ R & KBPEval2017 $\\parallel$ Trigger Identification $\\parallel$ F1 & KBPEval2017 $\\parallel$ Trigger Classification $\\parallel$ P & KBPEval2017 $\\parallel$ Trigger Classification $\\parallel$ R & KBPEval2017 $\\parallel$ Trigger Classification $\\parallel$ F1 \\\\\n\\midrule\nModel $\\parallel$ FBRNN(Char) & 61.3 & 45.6 & 52.3 & 57.5 & 42.8 & 49.1 & 57.97 & 36.92 & 45.11 & 51.71 & 32.94 & 40.24 \\\\\nModel $\\parallel$ DMCNN(Char) & 60.1 & 61.6 & 60.9 & 57.1 & 58.5 & 57.8 & 53.67 & 49.92 & 51.73 & 50.03 & 46.53 & 48.22 \\\\\nModel $\\parallel$ C-BiLSTM* & 65.6 & 66.7 & 66.1 & 60.0 & 60.9 & 60.4 & - & - & - & - & - & - \\\\\nModel $\\parallel$ FBRNN(Word) & 64.1 & 63.7 & 63.9 & 59.9 & 59.6 & 59.7 & 65.10 & 46.86 & 54.50 & 60.05 & 43.22 & 50.27 \\\\\nModel $\\parallel$ DMCNN(Word) & 66.6 & 63.6 & 65.1 & 61.6 & 58.8 & 60.2 & 60.43 & 51.64 & 55.69 & 54.81 & 46.84 & 50.51 \\\\\nModel $\\parallel$ HNN* & 74.2 & 63.1 & 68.2 & 77.1 & 53.1 & 63.0 & - & - & - & - & - & - \\\\\nModel $\\parallel$ Rich-C* & 62.2 & 71.9 & 66.7 & 58.9 & 68.1 & 63.2 & - & - & - & - & - & - \\\\\nModel $\\parallel$ KBP2017 Best* & - & - & - & - & - & - & 67.76 & 45.92 & 54.74 & 62.69 & 42.48 & 50.64 \\\\\nModel $\\parallel$ NPN(Concat) & 76.5 & 59.8 & 67.1 & 72.8 & 56.9 & 63.9 & 64.58 & 50.31 & 56.56 & 59.14 & 46.07 & 51.80 \\\\\nModel $\\parallel$ NPN(General) & 71.5 & 63.2 & 67.1 & 67.3 & 59.6 & 63.2 & 63.67 & 51.32 & 56.83 & 57.78 & 46.58 & 51.57 \\\\\nModel $\\parallel$ NPN(Task-specific) & 64.8 & 73.8 & 69.0 & 60.9 & 69.3 & 64.8 & 64.32 & 53.16 & 58.21 & 57.63 & 47.63 & 52.15 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Experiment results on ACE2005 and KBPEval2017. * indicates the result adapted from the original paper. For KBPEval2017, \u201cTrigger Identification\u201d corresponds to the \u201cSpan\u201d metric and \u201cTrigger Classification\u201d corresponds to the \u201cType\u201d metric reported in official evaluation.\n
\n \n ACE2005 || Trigger Identification || P\n \n ACE2005 || Trigger Identification || R\n \n ACE2005 || Trigger Identification || F1\n \n ACE2005 || Trigger Classification || P\n \n ACE2005 || Trigger Classification || R\n \n ACE2005 || Trigger Classification || F1\n \n KBPEval2017 || Trigger Identification || P\n \n KBPEval2017 || Trigger Identification || R\n \n KBPEval2017 || Trigger Identification || F1\n \n KBPEval2017 || Trigger Classification || P\n \n KBPEval2017 || Trigger Classification || R\n \n KBPEval2017 || Trigger Classification || F1\n
\n Model || FBRNN(Char)\n \n 61.3\n \n 45.6\n \n 52.3\n \n 57.5\n \n 42.8\n \n 49.1\n \n 57.97\n \n 36.92\n \n 45.11\n \n 51.71\n \n 32.94\n \n 40.24\n
\n Model || DMCNN(Char)\n \n 60.1\n \n 61.6\n \n 60.9\n \n 57.1\n \n 58.5\n \n 57.8\n \n 53.67\n \n 49.92\n \n 51.73\n \n 50.03\n \n 46.53\n \n 48.22\n
\n Model || C-BiLSTM*\n \n 65.6\n \n 66.7\n \n 66.1\n \n 60.0\n \n 60.9\n \n 60.4\n \n -\n \n -\n \n -\n \n -\n \n -\n \n -\n
\n Model || FBRNN(Word)\n \n 64.1\n \n 63.7\n \n 63.9\n \n 59.9\n \n 59.6\n \n 59.7\n \n 65.10\n \n 46.86\n \n 54.50\n \n 60.05\n \n 43.22\n \n 50.27\n
\n Model || DMCNN(Word)\n \n 66.6\n \n 63.6\n \n 65.1\n \n 61.6\n \n 58.8\n \n 60.2\n \n 60.43\n \n 51.64\n \n 55.69\n \n 54.81\n \n 46.84\n \n 50.51\n
\n Model || HNN*\n \n 74.2\n \n 63.1\n \n 68.2\n \n 77.1\n \n 53.1\n \n 63.0\n \n -\n \n -\n \n -\n \n -\n \n -\n \n -\n
\n Model || Rich-C*\n \n 62.2\n \n 71.9\n \n 66.7\n \n 58.9\n \n 68.1\n \n 63.2\n \n -\n \n -\n \n -\n \n -\n \n -\n \n -\n
\n Model || KBP2017 Best*\n \n -\n \n -\n \n -\n \n -\n \n -\n \n -\n \n 67.76\n \n 45.92\n \n 54.74\n \n 62.69\n \n 42.48\n \n 50.64\n
\n Model || NPN(Concat)\n \n 76.5\n \n 59.8\n \n 67.1\n \n 72.8\n \n 56.9\n \n 63.9\n \n 64.58\n \n 50.31\n \n 56.56\n \n 59.14\n \n 46.07\n \n 51.80\n
\n Model || NPN(General)\n \n 71.5\n \n 63.2\n \n 67.1\n \n 67.3\n \n 59.6\n \n 63.2\n \n 63.67\n \n 51.32\n \n 56.83\n \n 57.78\n \n 46.58\n \n 51.57\n
\n Model || NPN(Task-specific)\n \n 64.8\n \n 73.8\n \n 69.0\n \n 60.9\n \n 69.3\n \n 64.8\n \n 64.32\n \n 53.16\n \n 58.21\n \n 57.63\n \n 47.63\n \n 52.15\n
\n" }, { "instance_id": "5236b9f6-d423-4b5d-a8b5-1f1e21e38c5b", "table_id_paper": "P18-1168table_4", "caption": "Results on the development, public test (Test-P) and hidden test (Test-H) sets. For each model, we report both accuracy and consistency.", "row_header_level": 2, "row_headers": [ [ "Model", "MAJORITY" ], [ "Model", "MAXENT" ], [ "Model", "RULE" ], [ "Model", "SUP." ], [ "Model", "SUP.+DISC" ], [ "Model", "WEAKSUP." ], [ "Model", "W.+DISC" ] ], "column_header_level": 2, "column_headers": [ [ "Dev.", "Acc." ], [ "Dev.", "Con." ], [ "Test-P", "Acc." ], [ "Test-P", "Con." ], [ "Test-H", "Acc." ], [ "Test-H", "Con." ] ], "contents": [ [ "55.3", "-", "56.2", "-", "55.4", "-" ], [ "68.0", "-", "67.7", "-", "67.8", "-" ], [ "66.0", "29.2", "66.3", "32.7", "-", "-" ], [ "67.7", "36.7", "66.9", "38.3", "-", "-" ], [ "77.7", "52.4", "76.6", "51.8", "-", "-" ], [ "84.3", "66.3", "81.7", "60.1", "-", "-" ], [ "85.7", "67.4", "84.0", "65.0", "82.5", "63.9" ] ], "metrics_loc": "column", "metrics_type": "['Acc.', 'Con.', 'Acc.', 'Con.', 'Acc.', 'Con.']", "target_entity": "['W.+DISC']", "table_html": "
Dev. || Acc. Dev. || Con. Test-P || Acc. Test-P || Con. Test-H || Acc. Test-H || Con.
Model || MAJORITY 55.3 - 56.2 - 55.4 -
Model || MAXENT 68.0 - 67.7 - 67.8 -
Model || RULE 66.0 29.2 66.3 32.7 - -
Model || SUP. 67.7 36.7 66.9 38.3 - -
Model || SUP.+DISC 77.7 52.4 76.6 51.8 - -
Model || WEAKSUP. 84.3 66.3 81.7 60.1 - -
Model || W.+DISC 85.7 67.4 84.0 65.0 82.5 63.9
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P18-1168", "page_no": 8, "dir": "acl2018", "description": "Table 4 describes our main results. Our weakly-supervised semantic parser with re-ranking (W.+DISC) obtains 84.0 accuracy and 65.0 consistency on the public test set and 82.5 accuracy and 63.9 on the hidden one, improving accuracy by 14.7 points compared to state-of-theart. The accuracy of the rule-based parser (RULE) is less than 2 points below MAXENT, showing that a semantic parsing approach is very suitable for this task. The supervised parser obtains better performance (especially in consistency), and with re-ranking reaches 76.6 accuracy, showing that generalizing from generated examples is better than memorizing manually-defined patterns. Our weakly-supervised parser significantly improves over SUP., reaching an accuracy of 81.7 before reranking, and 84.0 after re-ranking (on the public test set). Consistency results show an even crisper trend of improvement across the models.", "class_sentence": "[1, 1, 1, 1, 1, 1]", "sentences": "['Table 4 describes our main results.', 'Our weakly-supervised semantic parser with re-ranking (W.+DISC) obtains 84.0 accuracy and 65.0 consistency on the public test set and 82.5 accuracy and 63.9 on the hidden one, improving accuracy by 14.7 points compared to state-of-theart.', 'The accuracy of the rule-based parser (RULE) is less than 2 points below MAXENT, showing that a semantic parsing approach is very suitable for this task.', 'The supervised parser obtains better performance (especially in consistency), and with re-ranking reaches 76.6 accuracy, showing that generalizing from generated examples is better than memorizing manually-defined patterns.', 'Our weakly-supervised parser significantly improves over SUP., reaching an accuracy of 81.7 before reranking, and 84.0 after re-ranking (on the public test set).', 'Consistency results show an even crisper trend of improvement across the models.']", "header_mention": "[None, ['W.+DISC', 'Test-P', 'Acc.', 'Con.', 'Test-H'], ['RULE', 'Acc.', 'MAXENT'], ['SUP.+DISC', 'Test-P', 'Acc.'], ['WEAKSUP.', 'SUP.', 'Test-P', 'Acc.'], ['Con.']]", "valid": 1, "image_id": "P18-1168-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results on the development, public test (Test-P) and hidden test (Test-H) sets. For each model, we report both accuracy and consistency.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Dev. || Acc.Dev. || Con.Test-P || Acc.Test-P || Con.Test-H || Acc.Test-H || Con.
Model || MAJORITY55.3-56.2-55.4-
Model || MAXENT68.0-67.7-67.8-
Model || RULE66.029.266.332.7--
Model || SUP.67.736.766.938.3--
Model || SUP.+DISC77.752.476.651.8--
Model || WEAKSUP.84.366.381.760.1--
Model || W.+DISC85.767.484.065.082.563.9
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results on the development, public test (Test-P) and hidden test (Test-H) sets. For each model, we report both accuracy and consistency.}\n\\begin{tabular}{lrlrlll}\n\\toprule\n & Dev. $\\parallel$ Acc. & Dev. $\\parallel$ Con. & Test-P $\\parallel$ Acc. & Test-P $\\parallel$ Con. & Test-H $\\parallel$ Acc. & Test-H $\\parallel$ Con. \\\\\n\\midrule\nModel $\\parallel$ MAJORITY & 55.3 & - & 56.2 & - & 55.4 & - \\\\\nModel $\\parallel$ MAXENT & 68 & - & 67.7 & - & 67.8 & - \\\\\nModel $\\parallel$ RULE & 66 & 29.2 & 66.3 & 32.7 & - & - \\\\\nModel $\\parallel$ SUP. & 67.7 & 36.7 & 66.9 & 38.3 & - & - \\\\\nModel $\\parallel$ SUP.+DISC & 77.7 & 52.4 & 76.6 & 51.8 & - & - \\\\\nModel $\\parallel$ WEAKSUP. & 84.3 & 66.3 & 81.7 & 60.1 & - & - \\\\\nModel $\\parallel$ W.+DISC & 85.7 & 67.4 & 84 & 65.0 & 82.5 & 63.9 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Results on the development, public test (Test-P) and hidden test (Test-H) sets. For each model, we report both accuracy and consistency.\n
\n \n Dev. || Acc.\n \n Dev. || Con.\n \n Test-P || Acc.\n \n Test-P || Con.\n \n Test-H || Acc.\n \n Test-H || Con.\n
\n Model || MAJORITY\n \n 55.3\n \n -\n \n 56.2\n \n -\n \n 55.4\n \n -\n
\n Model || MAXENT\n \n 68.0\n \n -\n \n 67.7\n \n -\n \n 67.8\n \n -\n
\n Model || RULE\n \n 66.0\n \n 29.2\n \n 66.3\n \n 32.7\n \n -\n \n -\n
\n Model || SUP.\n \n 67.7\n \n 36.7\n \n 66.9\n \n 38.3\n \n -\n \n -\n
\n Model || SUP.+DISC\n \n 77.7\n \n 52.4\n \n 76.6\n \n 51.8\n \n -\n \n -\n
\n Model || WEAKSUP.\n \n 84.3\n \n 66.3\n \n 81.7\n \n 60.1\n \n -\n \n -\n
\n Model || W.+DISC\n \n 85.7\n \n 67.4\n \n 84.0\n \n 65.0\n \n 82.5\n \n 63.9\n
\n" }, { "instance_id": "678a1d02-110a-4159-97ff-6d01ff8c45d2", "table_id_paper": "P18-1177table_5", "caption": "Human evaluation results for question generation. \u201cGrammaticality\u201d, \u201cMaking Sense\u201d and \u201cAnswerability\u201d are rated on a 1\u20135 scale (5 for the best, see the supplementary materials for a detailed rating scheme), \u201cAverage rank\u201d is rated on a 1\u20133 scale (1 for the most preferred, ties are allowed.) Two-tailed t-test results are shown for our method compared to ContextNQG (stat. significance is indicated with \u2217(p < 0.05), \u2217\u2217(p < 0.01).)", "row_header_level": 1, "row_headers": [ [ "ContextNQG" ], [ "CorefNQG" ], [ "Human" ] ], "column_header_level": 1, "column_headers": [ [ "Grammaticality" ], [ "Making Sense" ], [ "Answerability" ], [ "Avg. rank" ] ], "contents": [ [ "3.793", "3.836", "3.892", "1.768" ], [ "3.804*", "3.847**", "3.895*", "1.762" ], [ "3.807", "3.850", "3.902", "1.758" ] ], "metrics_loc": "column", "metrics_type": "['Grammaticality', 'Making Sense', 'Answerability', 'Avg. rank']", "target_entity": "['CorefNQG']", "table_html": "
Grammaticality Making Sense Answerability Avg. rank
ContextNQG 3.793 3.836 3.892 1.768
CorefNQG 3.804* 3.847** 3.895* 1.762
Human 3.807 3.850 3.902 1.758
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "P18-1177", "page_no": 8, "dir": "acl2018", "description": "Table 5 shows the results of the human evaluation. Bold indicates top scores. We see that the original human questions are preferred over the two NQG ssystems\u0081foutputs, which is understandable given the examples in Figure 3. The humangenerated questions make more sense and correspond better with the provided answers, particularly when they require information in the preceding context. How exactly to capture the preceding context so as to ask better and more diverse questions is an interesting future direction for research. In terms of grammaticality, however, the neural models do quite well, achieving very close to human performance. In addition, we see that our method (CorefNQG) performs statistically significantly better across all metrics in comparison to the baseline model (ContextNQG), which has access to the entire preceding context in the passage.", "class_sentence": "[1, 1, 1, 1, 2, 1, 1]", "sentences": "['Table 5 shows the results of the human evaluation.', 'Bold indicates top scores.', 'We see that the original human questions are preferred over the two NQG ssystems\\x81foutputs, which is understandable given the examples in Figure 3.', 'The humangenerated questions make more sense and correspond better with the provided answers, particularly when they require information in the preceding context.', 'How exactly to capture the preceding context so as to ask better and more diverse questions is an interesting future direction for research.', 'In terms of grammaticality, however, the neural models do quite well, achieving very close to human performance.', 'In addition, we see that our method (CorefNQG) performs statistically significantly better across all metrics in comparison to the baseline model (ContextNQG), which has access to the entire preceding context in the passage.']", "header_mention": "[None, None, ['Human', 'ContextNQG', 'CorefNQG'], ['Human'], None, ['Grammaticality', 'CorefNQG', 'Human'], ['CorefNQG', 'ContextNQG']]", "valid": 1, "image_id": "P18-1177-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Human evaluation results for question generation. \u201cGrammaticality\u201d, \u201cMaking Sense\u201d and \u201cAnswerability\u201d are rated on a 1\u20135 scale (5 for the best, see the supplementary materials for a detailed rating scheme), \u201cAverage rank\u201d is rated on a 1\u20133 scale (1 for the most preferred, ties are allowed.) Two-tailed t-test results are shown for our method compared to ContextNQG (stat. significance is indicated with \u2217(p < 0.05), \u2217\u2217(p < 0.01).)

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n GrammaticalityMaking SenseAnswerabilityAvg. rank
ContextNQG3.7933.8363.8921.768
CorefNQG3.804*3.847**3.895*1.762
Human3.8073.8503.9021.758
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Human evaluation results for question generation. \u201cGrammaticality\u201d, \u201cMaking Sense\u201d and \u201cAnswerability\u201d are rated on a 1\u20135 scale (5 for the best, see the supplementary materials for a detailed rating scheme), \u201cAverage rank\u201d is rated on a 1\u20133 scale (1 for the most preferred, ties are allowed.) Two-tailed t-test results are shown for our method compared to ContextNQG (stat. significance is indicated with $\\ast$(p $<$ 0.05), $\\ast$$\\ast$(p $<$ 0.01).)}\n\\begin{tabular}{llllr}\n\\toprule\n & Grammaticality & Making Sense & Answerability & Avg. rank \\\\\n\\midrule\nContextNQG & 3.793 & 3.836 & 3.892 & 1.768 \\\\\nCorefNQG & 3.804* & 3.847** & 3.895* & 1.762 \\\\\nHuman & 3.807 & 3.850 & 3.902 & 1.758 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Human evaluation results for question generation. \u201cGrammaticality\u201d, \u201cMaking Sense\u201d and \u201cAnswerability\u201d are rated on a 1\u20135 scale (5 for the best, see the supplementary materials for a detailed rating scheme), \u201cAverage rank\u201d is rated on a 1\u20133 scale (1 for the most preferred, ties are allowed.) Two-tailed t-test results are shown for our method compared to ContextNQG (stat. significance is indicated with \u2217(p < 0.05), \u2217\u2217(p < 0.01).)\n
\n \n Grammaticality\n \n Making Sense\n \n Answerability\n \n Avg. rank\n
\n ContextNQG\n \n 3.793\n \n 3.836\n \n 3.892\n \n 1.768\n
\n CorefNQG\n \n 3.804*\n \n 3.847**\n \n 3.895*\n \n 1.762\n
\n Human\n \n 3.807\n \n 3.850\n \n 3.902\n \n 1.758\n
\n" }, { "instance_id": "8f93f89d-b2a5-4a4a-a5c9-691d78fc5cb9", "table_id_paper": "P18-1188table_3", "caption": "Human evaluations: Ranking of various systems. Rank 1st is best and rank 4th, worst. Numbers show the percentage of times a system gets ranked at a certain position.", "row_header_level": 2, "row_headers": [ [ "Models", "LEAD" ], [ "Models", "POINTERNET" ], [ "Models", "XNET" ], [ "Models", "HUMAN" ] ], "column_header_level": 1, "column_headers": [ [ "1st" ], [ "2nd" ], [ "3rd" ], [ "4th" ] ], "contents": [ [ "0.15", "0.17", "0.47", "0.21" ], [ "0.16", "0.05", "0.31", "0.48" ], [ "0.28", "0.53", "0.15", "0.04" ], [ "0.41", "0.25", "0.07", "0.27" ] ], "metrics_loc": "column", "metrics_type": "['percentage', 'percentage', 'percentage', 'percentage']", "target_entity": "['XNET']", "table_html": "
1st 2nd 3rd 4th
Models || LEAD 0.15 0.17 0.47 0.21
Models || POINTERNET 0.16 0.05 0.31 0.48
Models || XNET 0.28 0.53 0.15 0.04
Models || HUMAN 0.41 0.25 0.07 0.27
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "P18-1188", "page_no": 6, "dir": "acl2018", "description": "The results of our human evaluation study are shown in Table 3. As one might imagine, HUMAN gets ranked 1st most of the time (41%). However, it is closely followed by XNET which ranked 1st 28% of the time. In comparison, POINTERNET and LEAD were mostly ranked at 3rd and 4th places. We also carried out pairwise comparisons between all models in Table 3 for their statistical significance using a one-way ANOVA with post-hoc Tukey HSD tests with (p < 0.01). It showed that XNET is significantly better than LEAD and POINTERNET, and it does not differ significantly from HUMAN. On the other hand, POINTERNET does not differ significantly from LEAD and it differs significantly from both XNET and HUMAN. The human evaluation results corroborates our empirical results in Table 1 and Table 2: XNET is better than LEAD and POINTERNET in producing informative and fluent summaries.", "class_sentence": "[1, 1, 1, 1, 2, 1, 1, 2]", "sentences": "['The results of our human evaluation study are shown in Table 3.', 'As one might imagine, HUMAN gets ranked 1st most of the time (41%).', 'However, it is closely followed by XNET which ranked 1st 28% of the time.', 'In comparison, POINTERNET and LEAD were mostly ranked at 3rd and 4th places.', 'We also carried out pairwise comparisons between all models in Table 3 for their statistical significance using a one-way ANOVA with post-hoc Tukey HSD tests with (p < 0.01).', 'It showed that XNET is significantly better than LEAD and POINTERNET, and it does not differ significantly from HUMAN.', 'On the other hand, POINTERNET does not differ significantly from LEAD and it differs significantly from both XNET and HUMAN.', 'The human evaluation results corroborates our empirical results in Table 1 and Table 2: XNET is better than LEAD and POINTERNET in producing informative and fluent summaries.']", "header_mention": "[None, ['HUMAN'], ['XNET'], ['LEAD', 'POINTERNET'], None, ['XNET', 'LEAD', 'POINTERNET', 'HUMAN'], ['LEAD', 'POINTERNET', 'XNET', 'HUMAN'], ['XNET', 'LEAD', 'POINTERNET']]", "valid": 1, "image_id": "P18-1188-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Human evaluations: Ranking of various systems. Rank 1st is best and rank 4th, worst. Numbers show the percentage of times a system gets ranked at a certain position.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n 1st2nd3rd4th
Models || LEAD0.150.170.470.21
Models || POINTERNET0.160.050.310.48
Models || XNET0.280.530.150.04
Models || HUMAN0.410.250.070.27
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Human evaluations: Ranking of various systems. Rank 1st is best and rank 4th, worst. Numbers show the percentage of times a system gets ranked at a certain position.}\n\\begin{tabular}{lrrrr}\n\\toprule\n & 1st & 2nd & 3rd & 4th \\\\\n\\midrule\nModels $\\parallel$ LEAD & 0.15 & 0.17 & 0.47 & 0.21 \\\\\nModels $\\parallel$ POINTERNET & 0.16 & 0.05 & 0.31 & 0.48 \\\\\nModels $\\parallel$ XNET & 0.28 & 0.53 & 0.15 & 0.04 \\\\\nModels $\\parallel$ HUMAN & 0.41 & 0.25 & 0.07 & 0.27 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Human evaluations: Ranking of various systems. Rank 1st is best and rank 4th, worst. Numbers show the percentage of times a system gets ranked at a certain position.\n
\n \n 1st\n \n 2nd\n \n 3rd\n \n 4th\n
\n Models || LEAD\n \n 0.15\n \n 0.17\n \n 0.47\n \n 0.21\n
\n Models || POINTERNET\n \n 0.16\n \n 0.05\n \n 0.31\n \n 0.48\n
\n Models || XNET\n \n 0.28\n \n 0.53\n \n 0.15\n \n 0.04\n
\n Models || HUMAN\n \n 0.41\n \n 0.25\n \n 0.07\n \n 0.27\n
\n" }, { "instance_id": "45e21b91-c3a0-4d23-8c4a-23014164b56a", "table_id_paper": "P18-1191table_6", "caption": "Joint span detection and question generation results on the dense development set, using exact-match for both spans and questions.", "row_header_level": 1, "row_headers": [ [ "Span + Local" ], [ "Span + Seq. (tau = 0.5)" ] ], "column_header_level": 1, "column_headers": [ [ "P" ], [ "R" ], [ "F" ] ], "contents": [ [ "37.8", "43.7", "40.6" ], [ "39.6", "45.8", "42.4" ] ], "metrics_loc": "column", "metrics_type": "['P', 'R', 'F']", "target_entity": "['Span + Seq. (tau = 0.5)']", "table_html": "
P R F
Span + Local 37.8 43.7 40.6
Span + Seq. (tau = 0.5) 39.6 45.8 42.4
", "table_name": "Table 6", "table_id": "table_6", "paper_id": "P18-1191", "page_no": 6, "dir": "acl2018", "description": "Table 6 shows precision and recall for joint span detection and question generation, using exact match for both. This metric is exceedingly hard, but it shows that almost 40% of predictions are exactly correct in both span and question.", "class_sentence": "[1, 1]", "sentences": "['Table 6 shows precision and recall for joint span detection and question generation, using exact match for both.', 'This metric is exceedingly hard, but it shows that almost 40% of predictions are exactly correct in both span and question.']", "header_mention": "[['Span + Local', 'Span + Seq. (tau = 0.5)', 'P', 'R'], ['Span + Seq. (tau = 0.5)', 'P', 'R']]", "valid": 1, "image_id": "P18-1191-Table6-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Joint span detection and question generation results on the dense development set, using exact-match for both spans and questions.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n PRF
Span + Local37.843.740.6
Span + Seq. (tau = 0.5)39.645.842.4
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{5}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Joint span detection and question generation results on the dense development set, using exact-match for both spans and questions.}\n\\begin{tabular}{lrrr}\n\\toprule\n & P & R & F \\\\\n\\midrule\nSpan + Local & 37.8 & 43.7 & 40.6 \\\\\nSpan + Seq. (tau $=$ 0.5) & 39.6 & 45.8 & 42.4 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 6: Joint span detection and question generation results on the dense development set, using exact-match for both spans and questions.\n
\n \n P\n \n R\n \n F\n
\n Span + Local\n \n 37.8\n \n 43.7\n \n 40.6\n
\n Span + Seq. (tau = 0.5)\n \n 39.6\n \n 45.8\n \n 42.4\n
\n" }, { "instance_id": "638d7451-b6f0-4887-8626-4b2b123c1271", "table_id_paper": "P18-1211table_2", "caption": "Performance of our approach on sentence ordering dataset from Barzilay and Lapata (2008).", "row_header_level": 2, "row_headers": [ [ "-", "Sequential CG" ], [ "-", "VLV-GM (2017)" ], [ "-", "HMM (2012)" ], [ "-", "HMM+Entity (2012)" ], [ "-", "HMM+Content (2012)" ], [ "Discriminative approaches", "DM (2017)" ], [ "Discriminative approaches", "Recursive (2014)" ], [ "Discriminative approaches", "Entity-Grid (2008)" ], [ "Discriminative approaches", "Graph (2013)" ] ], "column_header_level": 1, "column_headers": [ [ "Accidents" ], [ "Earthquakes" ] ], "contents": [ [ "0.813", "0.946" ], [ "0.77", "0.931" ], [ "0.822", "0.938" ], [ "0.842", "0.911" ], [ "0.742", "0.953" ], [ "0.93", "0.992" ], [ "0.864", "0.976" ], [ "0.904", "0.872" ], [ "0.846", "0.635" ] ], "metrics_loc": "column", "metrics_type": "['accuracy', 'accuracy']", "target_entity": "['Sequential CG']", "table_html": "
Accidents Earthquakes
- || Sequential CG 0.813 0.946
- || VLV-GM (2017) 0.77 0.931
- || HMM (2012) 0.822 0.938
- || HMM+Entity (2012) 0.842 0.911
- || HMM+Content (2012) 0.742 0.953
Discriminative approaches || DM (2017) 0.93 0.992
Discriminative approaches || Recursive (2014) 0.864 0.976
Discriminative approaches || Entity-Grid (2008) 0.904 0.872
Discriminative approaches || Graph (2013) 0.846 0.635
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P18-1211", "page_no": 8, "dir": "acl2018", "description": "Table 2 shows performance of the method compared with other approaches for coherence prediction. We note that Sequential-CG performs competitively with the state-of-the-art for generative approaches for the task, while needing no other annotation. In comparison, the HMM based approaches use significant annotation and syntactic features. Sequential-CG also outperforms several discriminative approaches for the task.", "class_sentence": "[1, 2, 1, 1]", "sentences": "['Table 2 shows performance of the method compared with other approaches for coherence prediction.', 'We note that Sequential-CG performs competitively with the state-of-the-art for generative approaches for the task, while needing no other annotation.', 'In comparison, the HMM based approaches use significant annotation and syntactic features.', 'Sequential-CG also outperforms several discriminative approaches for the task.']", "header_mention": "[None, ['Sequential CG'], ['HMM (2012)'], ['Sequential CG', 'Entity-Grid (2008)', 'Graph (2013)', 'Earthquakes']]", "valid": 1, "image_id": "P18-1211-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance of our approach on sentence ordering dataset from Barzilay and Lapata (2008).

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n AccidentsEarthquakes
- || Sequential CG0.8130.946
- || VLV-GM (2017)0.770.931
- || HMM (2012)0.8220.938
- || HMM+Entity (2012)0.8420.911
- || HMM+Content (2012)0.7420.953
Discriminative approaches || DM (2017)0.930.992
Discriminative approaches || Recursive (2014)0.8640.976
Discriminative approaches || Entity-Grid (2008)0.9040.872
Discriminative approaches || Graph (2013)0.8460.635
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance of our approach on sentence ordering dataset from Barzilay and Lapata (2008).}\n\\begin{tabular}{lrr}\n\\toprule\n & Accidents & Earthquakes \\\\\n\\midrule\n- $\\parallel$ Sequential CG & 0.813 & 0.946 \\\\\n- $\\parallel$ VLV-GM (2017) & 0.77 & 0.931 \\\\\n- $\\parallel$ HMM (2012) & 0.822 & 0.938 \\\\\n- $\\parallel$ HMM+Entity (2012) & 0.842 & 0.911 \\\\\n- $\\parallel$ HMM+Content (2012) & 0.742 & 0.953 \\\\\nDiscriminative approaches $\\parallel$ DM (2017) & 0.93 & 0.992 \\\\\nDiscriminative approaches $\\parallel$ Recursive (2014) & 0.864 & 0.976 \\\\\nDiscriminative approaches $\\parallel$ Entity-Grid (2008) & 0.904 & 0.872 \\\\\nDiscriminative approaches $\\parallel$ Graph (2013) & 0.846 & 0.635 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Performance of our approach on sentence ordering dataset from Barzilay and Lapata (2008).\n
\n \n Accidents\n \n Earthquakes\n
\n - || Sequential CG\n \n 0.813\n \n 0.946\n
\n - || VLV-GM (2017)\n \n 0.77\n \n 0.931\n
\n - || HMM (2012)\n \n 0.822\n \n 0.938\n
\n - || HMM+Entity (2012)\n \n 0.842\n \n 0.911\n
\n - || HMM+Content (2012)\n \n 0.742\n \n 0.953\n
\n Discriminative approaches || DM (2017)\n \n 0.93\n \n 0.992\n
\n Discriminative approaches || Recursive (2014)\n \n 0.864\n \n 0.976\n
\n Discriminative approaches || Entity-Grid (2008)\n \n 0.904\n \n 0.872\n
\n Discriminative approaches || Graph (2013)\n \n 0.846\n \n 0.635\n
\n" }, { "instance_id": "681b2274-6aa1-43a6-acd7-8ac42bca4efd", "table_id_paper": "P18-1230table_3", "caption": "F1-score (%) of different passes from 1 to 5 on the test data sets. It shows that appropriate number of passes can boost the performance as well as avoid over-fitting of the model.", "row_header_level": 2, "row_headers": [ [ "Pass", "1" ], [ "Pass", "2" ], [ "Pass", "3" ], [ "Pass", "4" ], [ "Pass", "5" ] ], "column_header_level": 1, "column_headers": [ [ "SE2" ], [ "SE3" ], [ "SE13" ], [ "SE15" ], [ "ALL" ] ], "contents": [ [ "71.6", "70.3", "67.0", "72.5", "70.3" ], [ "71.9", "70.2", "67.1", "72.8", "70.4" ], [ "72.2", "70.5", "67.2", "72.6", "70.6" ], [ "72.1", "70.4", "67.2", "72.4", "70.5" ], [ "72.0", "70.4", "67.1", "71.5", "70.3" ] ], "metrics_loc": "column", "metrics_type": "['F1-score', 'F1-score', 'F1-score', 'F1-score', 'F1-score']", "target_entity": "['Pass']", "table_html": "
SE2 SE3 SE13 SE15 ALL
Pass || 1 71.6 70.3 67.0 72.5 70.3
Pass || 2 71.9 70.2 67.1 72.8 70.4
Pass || 3 72.2 70.5 67.2 72.6 70.6
Pass || 4 72.1 70.4 67.2 72.4 70.5
Pass || 5 72.0 70.4 67.1 71.5 70.3
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "P18-1230", "page_no": 8, "dir": "acl2018", "description": "Furthermore, Table 3 shows the effectiveness of multi-pass operation in the memory module. It shows that multiple passes operation performs better than one pass, though the improvement is not significant. The reason of this phenomenon is that for most target words, one main word sense accounts for the majority of their appearances. Therefore, in most circumstances, one-pass inference can lead to the correct word senses. Case studies in Table 2 show that the proposed multipass inference can help to recognize the infrequent senses like the third sense for word play. In Table 3, with the increasing number of passes, the F1-score increases. However, when the number of passes is larger than 3, the F1-score stops increasing or even decreases due to over-fitting. It shows that appropriate number of passes can boost the performance as well as avoid over-fitting of the model.", "class_sentence": "[1, 1, 2, 2, 0, 1, 1, 2]", "sentences": "['Furthermore, Table 3 shows the effectiveness of multi-pass operation in the memory module.', 'It shows that multiple passes operation performs better than one pass, though the improvement is not significant.', 'The reason of this phenomenon is that for most target words, one main word sense accounts for the majority of their appearances.', 'Therefore, in most circumstances, one-pass inference can lead to the correct word senses.', 'Case studies in Table 2 show that the proposed multipass inference can help to recognize the infrequent senses like the third sense for word play.', 'In Table 3, with the increasing number of passes, the F1-score increases.', 'However, when the number of passes is larger than 3, the F1-score stops increasing or even decreases due to over-fitting.', 'It shows that appropriate number of passes can boost the performance as well as avoid over-fitting of the model.']", "header_mention": "[None, ['Pass', '1'], ['Pass', '1'], ['Pass', '1'], None, ['Pass', '1', '2', '3', 'SE2', 'SE3', 'SE13', 'SE15'], ['Pass', '4', '5'], ['Pass']]", "valid": 1, "image_id": "P18-1230-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

F1-score (%) of different passes from 1 to 5 on the test data sets. It shows that appropriate number of passes can boost the performance as well as avoid over-fitting of the model.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n SE2SE3SE13SE15ALL
Pass || 171.670.367.072.570.3
Pass || 271.970.267.172.870.4
Pass || 372.270.567.272.670.6
Pass || 472.170.467.272.470.5
Pass || 572.070.467.171.570.3
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{F1-score (\\%) of different passes from 1 to 5 on the test data sets. It shows that appropriate number of passes can boost the performance as well as avoid over-fitting of the model.}\n\\begin{tabular}{lrrrrr}\n\\toprule\n & SE2 & SE3 & SE13 & SE15 & ALL \\\\\n\\midrule\nPass $\\parallel$ 1 & 71.6 & 70.3 & 67 & 72.5 & 70.3 \\\\\nPass $\\parallel$ 2 & 71.9 & 70.2 & 67.1 & 72.8 & 70.4 \\\\\nPass $\\parallel$ 3 & 72.2 & 70.5 & 67.2 & 72.6 & 70.6 \\\\\nPass $\\parallel$ 4 & 72.1 & 70.4 & 67.2 & 72.4 & 70.5 \\\\\nPass $\\parallel$ 5 & 72 & 70.4 & 67.1 & 71.5 & 70.3 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: F1-score (%) of different passes from 1 to 5 on the test data sets. It shows that appropriate number of passes can boost the performance as well as avoid over-fitting of the model.\n
\n \n SE2\n \n SE3\n \n SE13\n \n SE15\n \n ALL\n
\n Pass || 1\n \n 71.6\n \n 70.3\n \n 67.0\n \n 72.5\n \n 70.3\n
\n Pass || 2\n \n 71.9\n \n 70.2\n \n 67.1\n \n 72.8\n \n 70.4\n
\n Pass || 3\n \n 72.2\n \n 70.5\n \n 67.2\n \n 72.6\n \n 70.6\n
\n Pass || 4\n \n 72.1\n \n 70.4\n \n 67.2\n \n 72.4\n \n 70.5\n
\n Pass || 5\n \n 72.0\n \n 70.4\n \n 67.1\n \n 71.5\n \n 70.3\n
\n" }, { "instance_id": "65730cde-1f94-4933-9ba3-552eafe0addb", "table_id_paper": "P18-1238table_2", "caption": "Human evaluation results on a sample from Conceptual Captions.", "row_header_level": 1, "row_headers": [ [ "Conceptual Captions" ] ], "column_header_level": 2, "column_headers": [ [ "GOOD (out of 3)", "1+" ], [ "GOOD (out of 3)", "2+" ], [ "GOOD (out of 3)", "3" ] ], "contents": [ [ "96.9%", "90.3%", "78.5%" ] ], "metrics_loc": "column", "metrics_type": "['GOOD (out of 3)', 'GOOD (out of 3)', 'GOOD (out of 3)']", "target_entity": "['Conceptual Captions']", "table_html": "
GOOD (out of 3) || 1+ GOOD (out of 3) || 2+ GOOD (out of 3) || 3
Conceptual Captions 96.9% 90.3% 78.5%
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P18-1238", "page_no": 6, "dir": "acl2018", "description": "The results are presented in Table 2 and show that, out of 3 annotations, over 90% of the captions receive a majority (2+) of GOOD judgments. This indicates that the Conceptual Captions pipeline, though involving extensive algorithmic processing, produces high-quality image captions.", "class_sentence": "[1, 2]", "sentences": "['The results are presented in Table 2 and show that, out of 3 annotations, over 90% of the captions receive a majority (2+) of GOOD judgments.', 'This indicates that the Conceptual Captions pipeline, though involving extensive algorithmic processing, produces high-quality image captions.']", "header_mention": "[['Conceptual Captions', 'GOOD (out of 3)', '1+', '2+'], ['Conceptual Captions']]", "valid": 1, "image_id": "P18-1238-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Human evaluation results on a sample from Conceptual Captions.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n GOOD (out of 3) || 1+GOOD (out of 3) || 2+GOOD (out of 3) || 3
Conceptual Captions96.9%90.3%78.5%
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Human evaluation results on a sample from Conceptual Captions.}\n\\begin{tabular}{llll}\n\\toprule\n & GOOD (out of 3) $\\parallel$ 1+ & GOOD (out of 3) $\\parallel$ 2+ & GOOD (out of 3) $\\parallel$ 3 \\\\\n\\midrule\nConceptual Captions & 96.9\\% & 90.3\\% & 78.5\\% \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Human evaluation results on a sample from Conceptual Captions.\n
\n \n GOOD (out of 3) || 1+\n \n GOOD (out of 3) || 2+\n \n GOOD (out of 3) || 3\n
\n Conceptual Captions\n \n 96.9%\n \n 90.3%\n \n 78.5%\n
\n" }, { "instance_id": "257afd2e-6cde-48b4-bb67-5f09401ace9e", "table_id_paper": "P18-1239table_1", "caption": "The proportion of images determined to be good representations of their corresponding word. In columns 2-5, we bucket the results by the word\u2019s ground-truth concreteness, while column 6 shows the results over all words. The last row shows the number of words in each bucket of concreteness, and the number of words overall for each language.", "row_header_level": 2, "row_headers": [ [ "Language", "English" ], [ "Language", "French" ], [ "Language", "Indonesian" ], [ "Language", "Uzbek" ], [ "Language", "All" ], [ "-", "# Words" ] ], "column_header_level": 2, "column_headers": [ [ "Concreteness Ratings", "1-2" ], [ "Concreteness Ratings", "2-3" ], [ "Concreteness Ratings", "3-4" ], [ "Concreteness Ratings", "4-5" ], [ "Concreteness Ratings", "Overall" ] ], "contents": [ [ "0.804", "0.814", "0.855", "0.913", "0.857" ], [ "0.622", "0.653", "0.706", "0.828", "0.721" ], [ "0.505", "0.569", "0.665", "0.785", "0.661" ], [ "0.568", "0.53", "0.594", "0.683", "0.601" ], [ "0.628", "0.649", "0.713", "0.81", "0.717" ], [ "77", "292", "292", "302", "963" ] ], "metrics_loc": "column", "metrics_type": "['Concreteness Ratings', 'Concreteness Ratings', 'Concreteness Ratings', 'Concreteness Ratings', 'Concreteness Ratings']", "target_entity": "['Concreteness Ratings']", "table_html": "
Concreteness Ratings || 1-2 Concreteness Ratings || 2-3 Concreteness Ratings || 3-4 Concreteness Ratings || 4-5 Concreteness Ratings || Overall
Language || English 0.804 0.814 0.855 0.913 0.857
Language || French 0.622 0.653 0.706 0.828 0.721
Language || Indonesian 0.505 0.569 0.665 0.785 0.661
Language || Uzbek 0.568 0.53 0.594 0.683 0.601
Language || All 0.628 0.649 0.713 0.81 0.717
- || # Words 77 292 292 302 963
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P18-1239", "page_no": 4, "dir": "acl2018", "description": "Table 1 shows the fraction of images that were judged to be good representations of the search word. It also demonstrates that as the concreteness of a word increases, the proportion of good images associated with that word increases as well. We further discuss the role of concreteness in Section 6.1. Overall, 85% of the English images, 72% of French, 66% of Indonesian, and 60% of Uzbek were judged to be good.", "class_sentence": "[1, 1, 0, 1]", "sentences": "['Table 1 shows the fraction of images that were judged to be good representations of the search word.', 'It also demonstrates that as the concreteness of a word increases, the proportion of good images associated with that word increases as well.', 'We further discuss the role of concreteness in Section 6.1.', 'Overall, 85% of the English images, 72% of French, 66% of Indonesian, and 60% of Uzbek were judged to be good.']", "header_mention": "[None, ['Concreteness Ratings'], None, ['Overall', 'English', 'French', 'Indonesian', 'Uzbek']]", "valid": 1, "image_id": "P18-1239-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

The proportion of images determined to be good representations of their corresponding word. In columns 2-5, we bucket the results by the word\u2019s ground-truth concreteness, while column 6 shows the results over all words. The last row shows the number of words in each bucket of concreteness, and the number of words overall for each language.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Concreteness Ratings || 1-2Concreteness Ratings || 2-3Concreteness Ratings || 3-4Concreteness Ratings || 4-5Concreteness Ratings || Overall
Language || English0.8040.8140.8550.9130.857
Language || French0.6220.6530.7060.8280.721
Language || Indonesian0.5050.5690.6650.7850.661
Language || Uzbek0.5680.530.5940.6830.601
Language || All0.6280.6490.7130.810.717
- || # Words77292292302963
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{The proportion of images determined to be good representations of their corresponding word. In columns 2-5, we bucket the results by the word\u2019s ground-truth concreteness, while column 6 shows the results over all words. The last row shows the number of words in each bucket of concreteness, and the number of words overall for each language.}\n\\begin{tabular}{lrrrrr}\n\\toprule\n & Concreteness Ratings $\\parallel$ 1-2 & Concreteness Ratings $\\parallel$ 2-3 & Concreteness Ratings $\\parallel$ 3-4 & Concreteness Ratings $\\parallel$ 4-5 & Concreteness Ratings $\\parallel$ Overall \\\\\n\\midrule\nLanguage $\\parallel$ English & 0.804 & 0.814 & 0.855 & 0.913 & 0.857 \\\\\nLanguage $\\parallel$ French & 0.622 & 0.653 & 0.706 & 0.828 & 0.721 \\\\\nLanguage $\\parallel$ Indonesian & 0.505 & 0.569 & 0.665 & 0.785 & 0.661 \\\\\nLanguage $\\parallel$ Uzbek & 0.568 & 0.53 & 0.594 & 0.683 & 0.601 \\\\\nLanguage $\\parallel$ All & 0.628 & 0.649 & 0.713 & 0.81 & 0.717 \\\\\n- $\\parallel$ \\# Words & 77 & 292 & 292 & 302 & 963 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: The proportion of images determined to be good representations of their corresponding word. In columns 2-5, we bucket the results by the word\u2019s ground-truth concreteness, while column 6 shows the results over all words. The last row shows the number of words in each bucket of concreteness, and the number of words overall for each language.\n
\n \n Concreteness Ratings || 1-2\n \n Concreteness Ratings || 2-3\n \n Concreteness Ratings || 3-4\n \n Concreteness Ratings || 4-5\n \n Concreteness Ratings || Overall\n
\n Language || English\n \n 0.804\n \n 0.814\n \n 0.855\n \n 0.913\n \n 0.857\n
\n Language || French\n \n 0.622\n \n 0.653\n \n 0.706\n \n 0.828\n \n 0.721\n
\n Language || Indonesian\n \n 0.505\n \n 0.569\n \n 0.665\n \n 0.785\n \n 0.661\n
\n Language || Uzbek\n \n 0.568\n \n 0.53\n \n 0.594\n \n 0.683\n \n 0.601\n
\n Language || All\n \n 0.628\n \n 0.649\n \n 0.713\n \n 0.81\n \n 0.717\n
\n - || # Words\n \n 77\n \n 292\n \n 292\n \n 302\n \n 963\n
\n" }, { "instance_id": "2bd92d63-cb35-45ea-8dc0-047973717dd0", "table_id_paper": "P18-2009table_3", "caption": "Results on the triplets data", "row_header_level": 2, "row_headers": [ [ "Method", "mean-vectors" ], [ "Method", "skip-thoughts-CS" ], [ "Method", "skip-thoughts-SICK" ], [ "Method", "triplets-sen" ] ], "column_header_level": 1, "column_headers": [ [ "accuracy" ] ], "contents": [ [ "0.65" ], [ "0.615" ], [ "0.547" ], [ "0.74" ] ], "metrics_loc": "column", "metrics_type": "['accuracy']", "target_entity": "['triplets-sen']", "table_html": "
accuracy
Method || mean-vectors 0.65
Method || skip-thoughts-CS 0.615
Method || skip-thoughts-SICK 0.547
Method || triplets-sen 0.74
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "P18-2009", "page_no": 4, "dir": "acl2018", "description": "Table 3 indicates that our method, denoted by triplet-sen, clearly outperforms the other tested methods. Surprizingly, skip-thoughts-SICK is inferior to skip-thoughts-CS.", "class_sentence": "[1, 1]", "sentences": "['Table 3 indicates that our method, denoted by triplet-sen, clearly outperforms the other tested methods.', 'Surprizingly, skip-thoughts-SICK is inferior to skip-thoughts-CS.']", "header_mention": "[['triplets-sen', 'mean-vectors', 'skip-thoughts-CS', 'skip-thoughts-SICK'], ['skip-thoughts-CS', 'skip-thoughts-SICK']]", "valid": 1, "image_id": "P18-2009-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results on the triplets data

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n accuracy
Method || mean-vectors0.65
Method || skip-thoughts-CS0.615
Method || skip-thoughts-SICK0.547
Method || triplets-sen0.74
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results on the triplets data}\n\\begin{tabular}{lr}\n\\toprule\n & accuracy \\\\\n\\midrule\nMethod $\\parallel$ mean-vectors & 0.65 \\\\\nMethod $\\parallel$ skip-thoughts-CS & 0.615 \\\\\nMethod $\\parallel$ skip-thoughts-SICK & 0.547 \\\\\nMethod $\\parallel$ triplets-sen & 0.74 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Results on the triplets data\n
\n \n accuracy\n
\n Method || mean-vectors\n \n 0.65\n
\n Method || skip-thoughts-CS\n \n 0.615\n
\n Method || skip-thoughts-SICK\n \n 0.547\n
\n Method || triplets-sen\n \n 0.74\n
\n" }, { "instance_id": "0a9e070c-0049-4b51-bece-ab951811e019", "table_id_paper": "P18-2015table_3", "caption": "Performance (AUCPR) of each noise reduction method; in bold are the best scores.", "row_header_level": 2, "row_headers": [ [ "System", "CNN+ONE" ], [ "System", "CNN+ATT" ], [ "System", "PCNN+ONE" ], [ "System", "PCNN+ATT" ] ], "column_header_level": 1, "column_headers": [ [ "Original" ], [ "Original+HITS" ], [ "Original+LSA" ], [ "Original+NMF" ], [ "Original+Ensemble" ] ], "contents": [ [ "0.18", "0.183", "0.173", "0.178", "0.181" ], [ "0.234", "0.235", "0.235", "0.233", "0.236" ], [ "0.231", "0.234", "0.233", "0.234", "0.235" ], [ "0.248", "0.253", "0.25", "0.252", "0.255" ] ], "metrics_loc": "column", "metrics_type": "['AUCPR', 'AUCPR', 'AUCPR', 'AUCPR', 'AUCPR']", "target_entity": "['Original+HITS', 'Original+Ensemble']", "table_html": "
Original Original+HITS Original+LSA Original+NMF Original+Ensemble
System || CNN+ONE 0.18 0.183 0.173 0.178 0.181
System || CNN+ATT 0.234 0.235 0.235 0.233 0.236
System || PCNN+ONE 0.231 0.234 0.233 0.234 0.235
System || PCNN+ATT 0.248 0.253 0.25 0.252 0.255
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "P18-2015", "page_no": 5, "dir": "acl2018", "description": "Table 3 presents the performance of noise reduction methods. Recall that the K-means-based method achieves a high P@50 for the seed selection method. Our assumption is that each cluster may represent a set in which elements have similar semantic properties. However, we observed that as the number of relations is relatively high and there is no distinct definition between some relations in the distantly labeled data (e.g., the following three relations /location/country/capital, are quite /location/province/capital, and /location/us state/capital, we decided not to perform the K-means-based method for our noise reduction task. The performances of the HITS-based, LSA-based, and NMF-based noise reduction methods are presented in Table 3. We experimentally set the portion of retained data from the distantly labeled data to 90%, given that the performance can be affected if too many sentences are removed from the original data. We also perform experiments with an ensemble method that combines the HITS-based and LSA-based strategies to merge rankings from their outputs, with half of the triples coming from the LSA-based method and the other half from the HITS-based method. Table 3 indicates that our proposed methods improved the performance of all CNN and PCNN models. Our ensemble method achieved the best improvements for three out of four systems, except that the HITS-based method obtained the best score for CNN+ONE.", "class_sentence": "[1, 0, 2, 2, 1, 2, 1, 1, 1]", "sentences": "['Table 3 presents the performance of noise reduction methods.', 'Recall that the K-means-based method achieves a high P@50 for the seed selection method.', 'Our assumption is that each cluster may represent a set in which elements have similar semantic properties.', 'However, we observed that as the number of relations is relatively high and there is no distinct definition between some relations in the distantly labeled data (e.g., the following three relations /location/country/capital, are quite /location/province/capital, and /location/us state/capital, we decided not to perform the K-means-based method for our noise reduction task.', 'The performances of the HITS-based, LSA-based, and NMF-based noise reduction methods are presented in Table 3.', 'We experimentally set the portion of retained data from the distantly labeled data to 90%, given that the performance can be affected if too many sentences are removed from the original data.', 'We also perform experiments with an ensemble method that combines the HITS-based and LSA-based strategies to merge rankings from their outputs, with half of the triples coming from the LSA-based method and the other half from the HITS-based method.', 'Table 3 indicates that our proposed methods improved the performance of all CNN and PCNN models.', 'Our ensemble method achieved the best improvements for three out of four systems, except that the HITS-based method obtained the best score for CNN+ONE.']", "header_mention": "[None, None, None, None, ['Original+HITS', 'Original+LSA', 'Original+NMF'], None, ['Original+Ensemble'], ['CNN+ONE', 'CNN+ATT', 'PCNN+ONE', 'PCNN+ATT', 'Original+HITS', 'Original+Ensemble'], ['Original+Ensemble', 'CNN+ATT', 'PCNN+ONE', 'PCNN+ATT']]", "valid": 1, "image_id": "P18-2015-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance (AUCPR) of each noise reduction method; in bold are the best scores.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n OriginalOriginal+HITSOriginal+LSAOriginal+NMFOriginal+Ensemble
System || CNN+ONE0.180.1830.1730.1780.181
System || CNN+ATT0.2340.2350.2350.2330.236
System || PCNN+ONE0.2310.2340.2330.2340.235
System || PCNN+ATT0.2480.2530.250.2520.255
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance (AUCPR) of each noise reduction method; in bold are the best scores.}\n\\begin{tabular}{lrrrrr}\n\\toprule\n & Original & Original+HITS & Original+LSA & Original+NMF & Original+Ensemble \\\\\n\\midrule\nSystem $\\parallel$ CNN+ONE & 0.18 & 0.183 & 0.173 & 0.178 & 0.181 \\\\\nSystem $\\parallel$ CNN+ATT & 0.234 & 0.235 & 0.235 & 0.233 & 0.236 \\\\\nSystem $\\parallel$ PCNN+ONE & 0.231 & 0.234 & 0.233 & 0.234 & 0.235 \\\\\nSystem $\\parallel$ PCNN+ATT & 0.248 & 0.253 & 0.25 & 0.252 & 0.255 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Performance (AUCPR) of each noise reduction method; in bold are the best scores.\n
\n \n Original\n \n Original+HITS\n \n Original+LSA\n \n Original+NMF\n \n Original+Ensemble\n
\n System || CNN+ONE\n \n 0.18\n \n 0.183\n \n 0.173\n \n 0.178\n \n 0.181\n
\n System || CNN+ATT\n \n 0.234\n \n 0.235\n \n 0.235\n \n 0.233\n \n 0.236\n
\n System || PCNN+ONE\n \n 0.231\n \n 0.234\n \n 0.233\n \n 0.234\n \n 0.235\n
\n System || PCNN+ATT\n \n 0.248\n \n 0.253\n \n 0.25\n \n 0.252\n \n 0.255\n
\n" }, { "instance_id": "2d2a02b6-b1fb-4eb4-9555-d04e3d90803d", "table_id_paper": "P18-2022table_1", "caption": "Average alignment performance across images. MSFC provides the best recall and lowest AER, and modified k-means the best precision. In all cases, the alignment framework yields stronger results than either of the timing-based baselines.", "row_header_level": 1, "row_headers": [ [ "Simultaneous" ], [ "1-second delay" ], [ "Alignment framework" ] ], "column_header_level": 2, "column_headers": [ [ "MSFC", "Precision" ], [ "MSFC", "Recall" ], [ "MSFC", "AER" ], [ "Modified k-means", "Precision" ], [ "Modified k-means", "Recall" ], [ "Modified k-means", "AER" ] ], "contents": [ [ "0.42", "0.30", "0.65", "0.49", "0.17", "0.74" ], [ "0.43", "0.31", "0.64", "0.50", "0.17", "0.74" ], [ "0.43", "0.50", "0.54", "0.56", "0.31", "0.60" ] ], "metrics_loc": "column", "metrics_type": "['Precision', 'Recall', 'AER', 'Precision', 'Recall', 'AER']", "target_entity": "['Alignment framework']", "table_html": "
MSFC || Precision MSFC || Recall MSFC || AER Modified k-means || Precision Modified k-means || Recall Modified k-means || AER
Simultaneous 0.42 0.30 0.65 0.49 0.17 0.74
1-second delay 0.43 0.31 0.64 0.50 0.17 0.74
Alignment framework 0.43 0.50 0.54 0.56 0.31 0.60
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P18-2022", "page_no": 5, "dir": "acl2018", "description": "Table 1 shows that the alignment framework performs better than either baseline. MSFC yields the highest recall and lowest AER with an absolute improvement of 0%, 19%, and 10% for precision, recall and AER, over the 1-second delay baseline. Modified k-means achieves higher precision with an absolute improvement of 6%, 14%, and 14%\nover baseline.", "class_sentence": "[1, 1, 1]", "sentences": "['Table 1 shows that the alignment framework performs better than either baseline.', 'MSFC yields the highest recall and lowest AER with an absolute improvement of 0%, 19%, and 10% for precision, recall and AER, over the 1-second delay baseline.', 'Modified k-means achieves higher precision with an absolute improvement of 6%, 14%, and 14%\\nover baseline.']", "header_mention": "[['Alignment framework', 'Simultaneous', '1-second delay'], ['Alignment framework', '1-second delay', 'MSFC', 'Precision', 'Recall', 'AER'], ['Alignment framework', 'Modified k-means', '1-second delay', 'Precision', 'Recall', 'AER']]", "valid": 1, "image_id": "P18-2022-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Average alignment performance across images. MSFC provides the best recall and lowest AER, and modified k-means the best precision. In all cases, the alignment framework yields stronger results than either of the timing-based baselines.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n MSFC || PrecisionMSFC || RecallMSFC || AERModified k-means || PrecisionModified k-means || RecallModified k-means || AER
Simultaneous0.420.300.650.490.170.74
1-second delay0.430.310.640.500.170.74
Alignment framework0.430.500.540.560.310.60
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Average alignment performance across images. MSFC provides the best recall and lowest AER, and modified k-means the best precision. In all cases, the alignment framework yields stronger results than either of the timing-based baselines.}\n\\begin{tabular}{lrrrrrr}\n\\toprule\n & MSFC $\\parallel$ Precision & MSFC $\\parallel$ Recall & MSFC $\\parallel$ AER & Modified k-means $\\parallel$ Precision & Modified k-means $\\parallel$ Recall & Modified k-means $\\parallel$ AER \\\\\n\\midrule\nSimultaneous & 0.42 & 0.3 & 0.65 & 0.49 & 0.17 & 0.74 \\\\\n1-second delay & 0.43 & 0.31 & 0.64 & 0.5 & 0.17 & 0.74 \\\\\nAlignment framework & 0.43 & 0.5 & 0.54 & 0.56 & 0.31 & 0.6 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Average alignment performance across images. MSFC provides the best recall and lowest AER, and modified k-means the best precision. In all cases, the alignment framework yields stronger results than either of the timing-based baselines.\n
\n \n MSFC || Precision\n \n MSFC || Recall\n \n MSFC || AER\n \n Modified k-means || Precision\n \n Modified k-means || Recall\n \n Modified k-means || AER\n
\n Simultaneous\n \n 0.42\n \n 0.30\n \n 0.65\n \n 0.49\n \n 0.17\n \n 0.74\n
\n 1-second delay\n \n 0.43\n \n 0.31\n \n 0.64\n \n 0.50\n \n 0.17\n \n 0.74\n
\n Alignment framework\n \n 0.43\n \n 0.50\n \n 0.54\n \n 0.56\n \n 0.31\n \n 0.60\n
\n" }, { "instance_id": "2d3c441c-3b28-4ba0-a8e1-3289d535be3c", "table_id_paper": "P19-1013table_8", "caption": "Results on math problems (\u00a75.5).", "row_header_level": 2, "row_headers": [ [ "Method", "depccg" ], [ "Method", "+ ELMo" ], [ "Method", "+ Proposed" ] ], "column_header_level": 1, "column_headers": [ [ "UF1" ], [ "LF1" ] ], "contents": [ [ "88.49", "66.15" ], [ "89.32", "70.74" ], [ "95.83", "80.53" ] ], "metrics_loc": "column", "metrics_type": "['UF1', 'LF1']", "target_entity": "['+ Proposed']", "table_html": "
UF1 LF1
Method || depccg 88.49 66.15
Method || + ELMo 89.32 70.74
Method || + Proposed 95.83 80.53
", "table_name": "Table 8", "table_id": "table_8", "paper_id": "P19-1013", "page_no": 9, "dir": "acl2019", "description": "Results . Table 8 shows the F1 scores of depccg in the respective settings. Remarkably, we observe huge additive performance improvement. While, in terms of labeled F1, ELMo contributes about 4 points on top of the plain depccg, adding the new training set (converted from dependency trees) improves more than 10 points. Examining the resulting trees, we observe that the huge gain is primarily involved with expressions unique to math.", "class_sentence": "[2, 1, 1, 1, 2]", "sentences": "['Results .', 'Table 8 shows the F1 scores of depccg in the respective settings.', 'Remarkably, we observe huge additive performance improvement.', 'While, in terms of labeled F1, ELMo contributes about 4 points on top of the plain depccg, adding the new training set (converted from dependency trees) improves more than 10 points.', 'Examining the resulting trees, we observe that the huge gain is primarily involved with expressions unique to math.']", "header_mention": "[None, ['UF1', 'LF1'], ['+ Proposed'], ['LF1', '+ ELMo', 'depccg', '+ Proposed'], None]", "valid": 1, "image_id": "P19-1013-Table8-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results on math problems (\u00a75.5).

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n UF1LF1
Method || depccg88.4966.15
Method || + ELMo89.3270.74
Method || + Proposed95.8380.53
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{7}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results on math problems ($\\S$5.5).}\n\\begin{tabular}{lrr}\n\\toprule\n & UF1 & LF1 \\\\\n\\midrule\nMethod $\\parallel$ depccg & 88.49 & 66.15 \\\\\nMethod $\\parallel$ + ELMo & 89.32 & 70.74 \\\\\nMethod $\\parallel$ + Proposed & 95.83 & 80.53 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 8: Results on math problems (\u00a75.5).\n
\n \n UF1\n \n LF1\n
\n Method || depccg\n \n 88.49\n \n 66.15\n
\n Method || + ELMo\n \n 89.32\n \n 70.74\n
\n Method || + Proposed\n \n 95.83\n \n 80.53\n
\n" }, { "instance_id": "cda79663-3b22-4a02-83a8-d64a1faa4390", "table_id_paper": "P19-1046table_4", "caption": "Unimodal, Bimodal and Trimodal Results of HFFN. Here, L, A and V denotes language, acoustic and visual modalities, respectively.", "row_header_level": 2, "row_headers": [ [ "Methods", "L" ], [ "Methods", "A" ], [ "Methods", "V" ], [ "Methods", "L+A" ], [ "Methods", "L+V" ], [ "Methods", "A+V" ], [ "Methods", "L+A+V" ] ], "column_header_level": 2, "column_headers": [ [ "CMU-MOSI", "Acc" ], [ "CMU-MOSI", "F1" ], [ "IEMOCAP", "Acc" ], [ "IEMOCAP", "F1" ] ], "contents": [ [ "78.59", "78.52", "81.46", "81.54" ], [ "48.14", "48.3", "38.08", "38.17" ], [ "56.97", "57.48", "34.52", "29.15" ], [ "78.06", "78.29", "80.38", "80.6" ], [ "79.39", "79.38", "80.05", "80.26" ], [ "55.17", "55.76", "55.17", "55.79" ], [ "80.19", "80.34", "82.37", "82.42" ] ], "metrics_loc": "column", "metrics_type": "['Acc', 'F1', 'Acc', 'F1']", "target_entity": "['L+A+V']", "table_html": "
CMU-MOSI || Acc CMU-MOSI || F1 IEMOCAP || Acc IEMOCAP || F1
Methods || L 78.59 78.52 81.46 81.54
Methods || A 48.14 48.3 38.08 38.17
Methods || V 56.97 57.48 34.52 29.15
Methods || L+A 78.06 78.29 80.38 80.6
Methods || L+V 79.39 79.38 80.05 80.26
Methods || A+V 55.17 55.76 55.17 55.79
Methods || L+A+V 80.19 80.34 82.37 82.42
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P19-1046", "page_no": 8, "dir": "acl2019", "description": "4.3.2 Discussion on Modality Importance. To explore the underlying information of each modality, we carry out an experiment to compare the performance among unimodal, bimodal and trimodal models. For unimodal models, we can infer from Table 4 that language modality is the most predictive for emotion prediction, outperforming acoustic and visual modalities with significant margin. When coupled with acoustic and visual modalities, the trimodal HFFN performs best, whose result is 1% ~ 2% better than the language-HFFN, indicting that acoustic and visual modalities actually play auxiliary roles while language is dominant. However, in our model, when conducting outer product, all three modalities are treated equally, which is probably not the optimal choice. In the future, we aim to develop a fusion technique paying more attention to the language modality, while the other two modalities only serve as accessory sources of information.", "class_sentence": "[2, 1, 1, 1, 1, 2]", "sentences": "['4.3.2 Discussion on Modality Importance.', 'To explore the underlying information of each modality, we carry out an experiment to compare the performance among unimodal, bimodal and trimodal models.', 'For unimodal models, we can infer from Table 4 that language modality is the most predictive for emotion prediction, outperforming acoustic and visual modalities with significant margin.', 'When coupled with acoustic and visual modalities, the trimodal HFFN performs best, whose result is 1% ~ 2% better than the language-HFFN, indicting that acoustic and visual modalities actually play auxiliary roles while language is dominant.', 'However, in our model, when conducting outer product, all three modalities are treated equally, which is probably not the optimal choice.', 'In the future, we aim to develop a fusion technique paying more attention to the language modality, while the other two modalities only serve as accessory sources of information.']", "header_mention": "[None, None, ['L', 'A', 'V'], ['L+A+V', 'L'], ['L+A+V'], None]", "valid": 1, "image_id": "P19-1046-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Unimodal, Bimodal and Trimodal Results of HFFN. Here, L, A and V denotes language, acoustic and visual modalities, respectively.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n CMU-MOSI || AccCMU-MOSI || F1IEMOCAP || AccIEMOCAP || F1
Methods || L78.5978.5281.4681.54
Methods || A48.1448.338.0838.17
Methods || V56.9757.4834.5229.15
Methods || L+A78.0678.2980.3880.6
Methods || L+V79.3979.3880.0580.26
Methods || A+V55.1755.7655.1755.79
Methods || L+A+V80.1980.3482.3782.42
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Unimodal, Bimodal and Trimodal Results of HFFN. Here, L, A and V denotes language, acoustic and visual modalities, respectively.}\n\\begin{tabular}{lrrrr}\n\\toprule\n & CMU-MOSI $\\parallel$ Acc & CMU-MOSI $\\parallel$ F1 & IEMOCAP $\\parallel$ Acc & IEMOCAP $\\parallel$ F1 \\\\\n\\midrule\nMethods $\\parallel$ L & 78.59 & 78.52 & 81.46 & 81.54 \\\\\nMethods $\\parallel$ A & 48.14 & 48.3 & 38.08 & 38.17 \\\\\nMethods $\\parallel$ V & 56.97 & 57.48 & 34.52 & 29.15 \\\\\nMethods $\\parallel$ L+A & 78.06 & 78.29 & 80.38 & 80.6 \\\\\nMethods $\\parallel$ L+V & 79.39 & 79.38 & 80.05 & 80.26 \\\\\nMethods $\\parallel$ A+V & 55.17 & 55.76 & 55.17 & 55.79 \\\\\nMethods $\\parallel$ L+A+V & 80.19 & 80.34 & 82.37 & 82.42 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Unimodal, Bimodal and Trimodal Results of HFFN. Here, L, A and V denotes language, acoustic and visual modalities, respectively.\n
\n \n CMU-MOSI || Acc\n \n CMU-MOSI || F1\n \n IEMOCAP || Acc\n \n IEMOCAP || F1\n
\n Methods || L\n \n 78.59\n \n 78.52\n \n 81.46\n \n 81.54\n
\n Methods || A\n \n 48.14\n \n 48.3\n \n 38.08\n \n 38.17\n
\n Methods || V\n \n 56.97\n \n 57.48\n \n 34.52\n \n 29.15\n
\n Methods || L+A\n \n 78.06\n \n 78.29\n \n 80.38\n \n 80.6\n
\n Methods || L+V\n \n 79.39\n \n 79.38\n \n 80.05\n \n 80.26\n
\n Methods || A+V\n \n 55.17\n \n 55.76\n \n 55.17\n \n 55.79\n
\n Methods || L+A+V\n \n 80.19\n \n 80.34\n \n 82.37\n \n 82.42\n
\n" }, { "instance_id": "ebd8237c-5909-42cb-a673-566d745042e1", "table_id_paper": "P19-1064table_2", "caption": "The overall mention detection results on the test set of OntoNotes. The F1 improvement is statistically significant under t-test with p < 0.05.", "row_header_level": 2, "row_headers": [ [ "Model", "Our full model" ], [ "Model", "Lee et al. (2018)" ] ], "column_header_level": 1, "column_headers": [ [ "Prec." ], [ "Rec." ], [ "F1" ] ], "contents": [ [ "89.6", "82.2", "85.7" ], [ "86.2", "83.7", "84.9" ] ], "metrics_loc": "column", "metrics_type": "['Prec.', 'Rec.', 'F1']", "target_entity": "None", "table_html": "
Prec. Rec. F1
Model || Our full model 89.6 82.2 85.7
Model || Lee et al. (2018) 86.2 83.7 84.9
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P19-1064", "page_no": 5, "dir": "acl2019", "description": "Since mention detection is a subtask of coreference resolution, it is worthwhile to study the performance. Table 2 shows the mention detection results on the test set. Similar to coreference linking results, our model achieves higher precision and F1 score, which indicates that our model can significantly reduce false positive mentions while it can still \u00ef\u00ac\u0081nd a reasonable number of mentions.", "class_sentence": "[2, 1, 1]", "sentences": "['Since mention detection is a subtask of coreference resolution, it is worthwhile to study the performance.', 'Table 2 shows the mention detection results on the test set.', 'Similar to coreference linking results, our model achieves higher precision and F1 score, which indicates that our model can significantly reduce false positive mentions while it can still \u00ef\u00ac\\x81nd a reasonable number of mentions.']", "header_mention": "[None, None, ['Our full model', 'Prec.', 'F1']]", "valid": 1, "image_id": "P19-1064-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

The overall mention detection results on the test set of OntoNotes. The F1 improvement is statistically significant under t-test with p < 0.05.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Prec.Rec.F1
Model || Our full model89.682.285.7
Model || Lee et al. (2018)86.283.784.9
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{The overall mention detection results on the test set of OntoNotes. The F1 improvement is statistically significant under t-test with p $<$ 0.05.}\n\\begin{tabular}{lrrr}\n\\toprule\n & Prec. & Rec. & F1 \\\\\n\\midrule\nModel $\\parallel$ Our full model & 89.6 & 82.2 & 85.7 \\\\\nModel $\\parallel$ Lee et al. (2018) & 86.2 & 83.7 & 84.9 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: The overall mention detection results on the test set of OntoNotes. The F1 improvement is statistically significant under t-test with p < 0.05.\n
\n \n Prec.\n \n Rec.\n \n F1\n
\n Model || Our full model\n \n 89.6\n \n 82.2\n \n 85.7\n
\n Model || Lee et al. (2018)\n \n 86.2\n \n 83.7\n \n 84.9\n
\n" }, { "instance_id": "84a21bff-1c9e-4a84-9e06-1679be731331", "table_id_paper": "P19-1117table_2", "caption": "Translation performance on one-to-two, one-to-three and one-to-four translation tasks. Rep denotes our proposed representor. Emb, Attn, and Dis represent our proposed language-sensitive methods to address multilingual translation. Note that the source language of all our experiments is English.", "row_header_level": 2, "row_headers": [ [ "One-to-Two", "De" ], [ "One-to-Two", "Lv" ], [ "One-to-Two", "De" ], [ "One-to-Two", "Fi" ], [ "One-to-Two", "De" ], [ "One-to-Two", "Zh" ], [ "One-to-Three", "De" ], [ "One-to-Three", "Zh" ], [ "One-to-Three", "Fi" ], [ "One-to-Three", "De" ], [ "One-to-Three", "Lv" ], [ "One-to-Three", "Fi" ], [ "One-to-Four", "De" ], [ "One-to-Four", "Lv" ], [ "One-to-Four", "Zh" ], [ "One-to-Four", "Fi" ] ], "column_header_level": 1, "column_headers": [ [ "NMT Baselines" ], [ "Multi-NMT Baselines Johnson et al.(2017)" ], [ "Three-Stgy Wang et al.(2018)" ], [ "Rep+Emb" ], [ "Rep+Emb+Attn" ], [ "Rep+Emb+Attn+Dis" ] ], "contents": [ [ "27.5", "27.26", "27.35", "26.6", "26.96", "27.74" ], [ "16.28", "16.32", "16.38", "15.37", "15.87", "16.79" ], [ "27.5", "27.88", "27.89", "26.96", "27.32", "27.96" ], [ "16.83", "16.47", "16.7", "15.78", "16.58", "16.89" ], [ "27.5", "26.8", "26.99", "26.08", "26.68", "27.45" ], [ "26.04", "25.54", "25.78", "24.48", "25.33", "26.17" ], [ "27.5", "25.44", "25.55", "24.82", "25.45", "26.06" ], [ "26.04", "24.87", "25.63", "24.12", "24.93", "26.12" ], [ "16.83", "16.86", "16.97", "16.06", "16.78", "17.12" ], [ "27.5", "25.98", "26.12", "24.88", "25.8", "26.42" ], [ "16.28", "14.88", "15.44", "14.51", "15.58", "16.31" ], [ "16.83", "16.94", "17.05", "16.15", "16.79", "17.22" ], [ "27.5", "23.59", "22.88", "22.88", "23.58", "24.08" ], [ "16.28", "15.57", "16.02", "15", "16.21", "16.57" ], [ "26.04", "25.24", "25.83", "24.15", "25.27", "26.29" ], [ "16.83", "13.45", "14.12", "12.99", "14.11", "15.03" ] ], "metrics_loc": "column", "metrics_type": "['BLEU', 'BLEU', 'BLEU', 'BLEU', 'BLEU', 'BLEU']", "target_entity": "['Rep+Emb+Attn+Dis']", "table_html": "
NMT Baselines Multi-NMT Baselines Johnson et al.(2017) Three-Stgy Wang et al.(2018) Rep+Emb Rep+Emb+Attn Rep+Emb+Attn+Dis
One-to-Two || De 27.5 27.26 27.35 26.6 26.96 27.74
One-to-Two || Lv 16.28 16.32 16.38 15.37 15.87 16.79
One-to-Two || De 27.5 27.88 27.89 26.96 27.32 27.96
One-to-Two || Fi 16.83 16.47 16.7 15.78 16.58 16.89
One-to-Two || De 27.5 26.8 26.99 26.08 26.68 27.45
One-to-Two || Zh 26.04 25.54 25.78 24.48 25.33 26.17
One-to-Three || De 27.5 25.44 25.55 24.82 25.45 26.06
One-to-Three || Zh 26.04 24.87 25.63 24.12 24.93 26.12
One-to-Three || Fi 16.83 16.86 16.97 16.06 16.78 17.12
One-to-Three || De 27.5 25.98 26.12 24.88 25.8 26.42
One-to-Three || Lv 16.28 14.88 15.44 14.51 15.58 16.31
One-to-Three || Fi 16.83 16.94 17.05 16.15 16.79 17.22
One-to-Four || De 27.5 23.59 22.88 22.88 23.58 24.08
One-to-Four || Lv 16.28 15.57 16.02 15 16.21 16.57
One-to-Four || Zh 26.04 25.24 25.83 24.15 25.27 26.29
One-to-Four || Fi 16.83 13.45 14.12 12.99 14.11 15.03
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P19-1117", "page_no": 6, "dir": "acl2019", "description": "The main results on the one-to-many translation scenario, including one-to-two, one-to-three and one-to-four translation tasks are reported in Table 2. We present a typical Multi-NMT adopting Johnson et al. (2017) method on Transformer as our Multi-NMT baselines model. Obviously, Multi-NMT Baselines cannot outperform NMT Baselines in all cases, among which four directions are comparable and twelve are worse. With respect to our proposed method, it is clear that our compact method consistently outperforms the baseline systems. Compared with another strong one-to-many translation model Three-Stgy proposed by Wang et al. (2018), our compact method can achieve better results as well. Moreover, our method can perform even better than individually trained systems in most cases (eleven out of sixteen cases). The results demonstrate the effectiveness of our method.", "class_sentence": "[1, 1, 1, 1, 1, 1, 1]", "sentences": "['The main results on the one-to-many translation scenario, including one-to-two, one-to-three and one-to-four translation tasks are reported in Table 2.', 'We present a typical Multi-NMT adopting Johnson et al. (2017) method on Transformer as our Multi-NMT baselines model.', 'Obviously, Multi-NMT Baselines cannot outperform NMT Baselines in all cases, among which four directions are comparable and twelve are worse.', 'With respect to our proposed method, it is clear that our compact method consistently outperforms the baseline systems.', 'Compared with another strong one-to-many translation model Three-Stgy proposed by Wang et al. (2018), our compact method can achieve better results as well.', 'Moreover, our method can perform even better than individually trained systems in most cases (eleven out of sixteen cases).', 'The results demonstrate the effectiveness of our method.']", "header_mention": "[['One-to-Two', 'One-to-Three', 'One-to-Four'], ['Multi-NMT Baselines Johnson et al.(2017)'], ['Multi-NMT Baselines Johnson et al.(2017)', 'NMT Baselines'], ['Rep+Emb+Attn+Dis'], ['Three-Stgy Wang et al.(2018)', 'Rep+Emb+Attn+Dis'], ['Rep+Emb+Attn+Dis'], ['Rep+Emb+Attn+Dis']]", "valid": 1, "image_id": "P19-1117-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Translation performance on one-to-two, one-to-three and one-to-four translation tasks. Rep denotes our proposed representor. Emb, Attn, and Dis represent our proposed language-sensitive methods to address multilingual translation. Note that the source language of all our experiments is English.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n NMT BaselinesMulti-NMT Baselines Johnson et al.(2017)Three-Stgy Wang et al.(2018)Rep+EmbRep+Emb+AttnRep+Emb+Attn+Dis
One-to-Two || De27.527.2627.3526.626.9627.74
One-to-Two || Lv16.2816.3216.3815.3715.8716.79
One-to-Two || De27.527.8827.8926.9627.3227.96
One-to-Two || Fi16.8316.4716.715.7816.5816.89
One-to-Two || De27.526.826.9926.0826.6827.45
One-to-Two || Zh26.0425.5425.7824.4825.3326.17
One-to-Three || De27.525.4425.5524.8225.4526.06
One-to-Three || Zh26.0424.8725.6324.1224.9326.12
One-to-Three || Fi16.8316.8616.9716.0616.7817.12
One-to-Three || De27.525.9826.1224.8825.826.42
One-to-Three || Lv16.2814.8815.4414.5115.5816.31
One-to-Three || Fi16.8316.9417.0516.1516.7917.22
One-to-Four || De27.523.5922.8822.8823.5824.08
One-to-Four || Lv16.2815.5716.021516.2116.57
One-to-Four || Zh26.0425.2425.8324.1525.2726.29
One-to-Four || Fi16.8313.4514.1212.9914.1115.03
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Translation performance on one-to-two, one-to-three and one-to-four translation tasks. Rep denotes our proposed representor. Emb, Attn, and Dis represent our proposed language-sensitive methods to address multilingual translation. Note that the source language of all our experiments is English.}\n\\begin{tabular}{lrrrrrr}\n\\toprule\n & NMT Baselines & Multi-NMT Baselines Johnson et al.(2017) & Three-Stgy Wang et al.(2018) & Rep+Emb & Rep+Emb+Attn & Rep+Emb+Attn+Dis \\\\\n\\midrule\nOne-to-Two $\\parallel$ De & 27.5 & 27.26 & 27.35 & 26.6 & 26.96 & 27.74 \\\\\nOne-to-Two $\\parallel$ Lv & 16.28 & 16.32 & 16.38 & 15.37 & 15.87 & 16.79 \\\\\nOne-to-Two $\\parallel$ De & 27.5 & 27.88 & 27.89 & 26.96 & 27.32 & 27.96 \\\\\nOne-to-Two $\\parallel$ Fi & 16.83 & 16.47 & 16.7 & 15.78 & 16.58 & 16.89 \\\\\nOne-to-Two $\\parallel$ De & 27.5 & 26.8 & 26.99 & 26.08 & 26.68 & 27.45 \\\\\nOne-to-Two $\\parallel$ Zh & 26.04 & 25.54 & 25.78 & 24.48 & 25.33 & 26.17 \\\\\nOne-to-Three $\\parallel$ De & 27.5 & 25.44 & 25.55 & 24.82 & 25.45 & 26.06 \\\\\nOne-to-Three $\\parallel$ Zh & 26.04 & 24.87 & 25.63 & 24.12 & 24.93 & 26.12 \\\\\nOne-to-Three $\\parallel$ Fi & 16.83 & 16.86 & 16.97 & 16.06 & 16.78 & 17.12 \\\\\nOne-to-Three $\\parallel$ De & 27.5 & 25.98 & 26.12 & 24.88 & 25.8 & 26.42 \\\\\nOne-to-Three $\\parallel$ Lv & 16.28 & 14.88 & 15.44 & 14.51 & 15.58 & 16.31 \\\\\nOne-to-Three $\\parallel$ Fi & 16.83 & 16.94 & 17.05 & 16.15 & 16.79 & 17.22 \\\\\nOne-to-Four $\\parallel$ De & 27.5 & 23.59 & 22.88 & 22.88 & 23.58 & 24.08 \\\\\nOne-to-Four $\\parallel$ Lv & 16.28 & 15.57 & 16.02 & 15 & 16.21 & 16.57 \\\\\nOne-to-Four $\\parallel$ Zh & 26.04 & 25.24 & 25.83 & 24.15 & 25.27 & 26.29 \\\\\nOne-to-Four $\\parallel$ Fi & 16.83 & 13.45 & 14.12 & 12.99 & 14.11 & 15.03 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Translation performance on one-to-two, one-to-three and one-to-four translation tasks. Rep denotes our proposed representor. Emb, Attn, and Dis represent our proposed language-sensitive methods to address multilingual translation. Note that the source language of all our experiments is English.\n
\n \n NMT Baselines\n \n Multi-NMT Baselines Johnson et al.(2017)\n \n Three-Stgy Wang et al.(2018)\n \n Rep+Emb\n \n Rep+Emb+Attn\n \n Rep+Emb+Attn+Dis\n
\n One-to-Two || De\n \n 27.5\n \n 27.26\n \n 27.35\n \n 26.6\n \n 26.96\n \n 27.74\n
\n One-to-Two || Lv\n \n 16.28\n \n 16.32\n \n 16.38\n \n 15.37\n \n 15.87\n \n 16.79\n
\n One-to-Two || De\n \n 27.5\n \n 27.88\n \n 27.89\n \n 26.96\n \n 27.32\n \n 27.96\n
\n One-to-Two || Fi\n \n 16.83\n \n 16.47\n \n 16.7\n \n 15.78\n \n 16.58\n \n 16.89\n
\n One-to-Two || De\n \n 27.5\n \n 26.8\n \n 26.99\n \n 26.08\n \n 26.68\n \n 27.45\n
\n One-to-Two || Zh\n \n 26.04\n \n 25.54\n \n 25.78\n \n 24.48\n \n 25.33\n \n 26.17\n
\n One-to-Three || De\n \n 27.5\n \n 25.44\n \n 25.55\n \n 24.82\n \n 25.45\n \n 26.06\n
\n One-to-Three || Zh\n \n 26.04\n \n 24.87\n \n 25.63\n \n 24.12\n \n 24.93\n \n 26.12\n
\n One-to-Three || Fi\n \n 16.83\n \n 16.86\n \n 16.97\n \n 16.06\n \n 16.78\n \n 17.12\n
\n One-to-Three || De\n \n 27.5\n \n 25.98\n \n 26.12\n \n 24.88\n \n 25.8\n \n 26.42\n
\n One-to-Three || Lv\n \n 16.28\n \n 14.88\n \n 15.44\n \n 14.51\n \n 15.58\n \n 16.31\n
\n One-to-Three || Fi\n \n 16.83\n \n 16.94\n \n 17.05\n \n 16.15\n \n 16.79\n \n 17.22\n
\n One-to-Four || De\n \n 27.5\n \n 23.59\n \n 22.88\n \n 22.88\n \n 23.58\n \n 24.08\n
\n One-to-Four || Lv\n \n 16.28\n \n 15.57\n \n 16.02\n \n 15\n \n 16.21\n \n 16.57\n
\n One-to-Four || Zh\n \n 26.04\n \n 25.24\n \n 25.83\n \n 24.15\n \n 25.27\n \n 26.29\n
\n One-to-Four || Fi\n \n 16.83\n \n 13.45\n \n 14.12\n \n 12.99\n \n 14.11\n \n 15.03\n
\n" }, { "instance_id": "13c44ae3-f01a-4420-ac11-3bbd4d5a2138", "table_id_paper": "P19-1134table_1", "caption": "Precision evaluated automatically for the top rated relation instances. \u2020 marks results reported in the original paper. \u2021 marks our results using the OpenNRE implementation.", "row_header_level": 2, "row_headers": [ [ "System", "Mintz" ], [ "System", "PCNN+ATT" ], [ "System", "RESIDE" ], [ "System", "DISTRE" ] ], "column_header_level": 1, "column_headers": [ [ "AUC" ], [ "P@100" ], [ "P@200" ], [ "P@300" ], [ "P@500" ], [ "P@1000" ], [ "P@2000" ] ], "contents": [ [ "0.107", "52.3", "50.2", "45", "39.7", "33.6", "23.4" ], [ "0.341", "73", "68", "67.3", "63.6", "53.3", "40" ], [ "0.415", "81.8", "75.4", "74.3", "69.7", "59.3", "45" ], [ "0.422", "68", "67", "65.3", "65", "60.2", "47.9" ] ], "metrics_loc": "column", "metrics_type": "['AUC', 'P@100', 'P@200', 'P@300', 'P@500', 'P@1000', 'P@2000']", "target_entity": "['DISTRE']", "table_html": "
AUC P@100 P@200 P@300 P@500 P@1000 P@2000
System || Mintz 0.107 52.3 50.2 45 39.7 33.6 23.4
System || PCNN+ATT 0.341 73 68 67.3 63.6 53.3 40
System || RESIDE 0.415 81.8 75.4 74.3 69.7 59.3 45
System || DISTRE 0.422 68 67 65.3 65 60.2 47.9
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P19-1134", "page_no": 6, "dir": "acl2019", "description": "5.1 Held-out Evaluation. DISTRE with selective attention achieves a new state-of-the-art AUC value of 0.422. The precision-recall curve in Figure 4 shows that it outperforms RESIDE and PCNN+ATT at higher recall levels, while precision is lower for top predicted relation instances. The results of the PCNN+ATT model indicate that its performance is only better in the very beginning of the curve, but its precision drops early and only achieves an AUC value of 0.341. Similar, RESIDE performs better in the beginning but drops in precision after a recall-level of approximately 0.25. This suggests that our method yields a more balanced overall performance, which we believe is important in many real-world applications. Table 1 also shows detailed precision values measured at different points along the P-R curve. We again can observe that while DISTRE has lower precision for the top 500 predicted relation instances, it shows a state-of-the-art precision of 60.2% for the top 1000 and continues to perform higher for the remaining, much larger part of the predictions.", "class_sentence": "[2, 1, 1, 1, 1, 1, 1, 1]", "sentences": "['5.1 Held-out Evaluation.', 'DISTRE with selective attention achieves a new state-of-the-art AUC value of 0.422.', 'The precision-recall curve in Figure 4 shows that it outperforms RESIDE and PCNN+ATT at higher recall levels, while precision is lower for top predicted relation instances.', 'The results of the PCNN+ATT model indicate that its performance is only better in the very beginning of the curve, but its precision drops early and only achieves an AUC value of 0.341.', 'Similar, RESIDE performs better in the beginning but drops in precision after a recall-level of approximately 0.25.', 'This suggests that our method yields a more balanced overall performance, which we believe is important in many real-world applications.', 'Table 1 also shows detailed precision values measured at different points along the P-R curve.', 'We again can observe that while DISTRE has lower precision for the top 500 predicted relation instances, it shows a state-of-the-art precision of 60.2% for the top 1000 and continues to perform higher for the remaining, much larger part of the predictions.']", "header_mention": "[None, ['DISTRE', 'AUC'], ['RESIDE', 'PCNN+ATT'], ['PCNN+ATT', 'AUC'], ['RESIDE'], ['DISTRE'], ['P@100', 'P@200', 'P@300', 'P@500', 'P@1000', 'P@2000'], ['DISTRE']]", "valid": 1, "image_id": "P19-1134-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Precision evaluated automatically for the top rated relation instances. \u2020 marks results reported in the original paper. \u2021 marks our results using the OpenNRE implementation.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n AUCP@100P@200P@300P@500P@1000P@2000
System || Mintz0.10752.350.24539.733.623.4
System || PCNN+ATT0.341736867.363.653.340
System || RESIDE0.41581.875.474.369.759.345
System || DISTRE0.422686765.36560.247.9
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Precision evaluated automatically for the top rated relation instances. $\\dagger$ marks results reported in the original paper. $\\ddagger$ marks our results using the OpenNRE implementation.}\n\\begin{tabular}{lrrrrrrr}\n\\toprule\n & AUC & P@100 & P@200 & P@300 & P@500 & P@1000 & P@2000 \\\\\n\\midrule\nSystem $\\parallel$ Mintz & 0.107 & 52.3 & 50.2 & 45 & 39.7 & 33.6 & 23.4 \\\\\nSystem $\\parallel$ PCNN+ATT & 0.341 & 73 & 68 & 67.3 & 63.6 & 53.3 & 40 \\\\\nSystem $\\parallel$ RESIDE & 0.415 & 81.8 & 75.4 & 74.3 & 69.7 & 59.3 & 45 \\\\\nSystem $\\parallel$ DISTRE & 0.422 & 68 & 67 & 65.3 & 65 & 60.2 & 47.9 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Precision evaluated automatically for the top rated relation instances. \u2020 marks results reported in the original paper. \u2021 marks our results using the OpenNRE implementation.\n
\n \n AUC\n \n P@100\n \n P@200\n \n P@300\n \n P@500\n \n P@1000\n \n P@2000\n
\n System || Mintz\n \n 0.107\n \n 52.3\n \n 50.2\n \n 45\n \n 39.7\n \n 33.6\n \n 23.4\n
\n System || PCNN+ATT\n \n 0.341\n \n 73\n \n 68\n \n 67.3\n \n 63.6\n \n 53.3\n \n 40\n
\n System || RESIDE\n \n 0.415\n \n 81.8\n \n 75.4\n \n 74.3\n \n 69.7\n \n 59.3\n \n 45\n
\n System || DISTRE\n \n 0.422\n \n 68\n \n 67\n \n 65.3\n \n 65\n \n 60.2\n \n 47.9\n
\n" }, { "instance_id": "8fac0cee-5356-4ed6-bad7-ef64cd0c7573", "table_id_paper": "P19-1170table_2", "caption": "Comparisons without post-processing of methods. Results reproduced from (Smith et al., 2017) for fair comparison. Left: Comparisons using the same expert dictionary as (Smith et al., 2017). Right: Comparisons using the pseudo-dictionary from (Smith et al., 2017).", "row_header_level": 2, "row_headers": [ [ "Methods", "Mikolov et. al." ], [ "Methods", "CCA (Sklearn)" ], [ "Methods", "CCA" ], [ "Methods", "SVD" ], [ "Methods", "IBFA (Ours)" ] ], "column_header_level": 2, "column_headers": [ [ "English to Italian", "@1" ], [ "English to Italian", "@5" ], [ "English to Italian", "@10" ], [ "Italian to English", "@1" ], [ "Italian to English", "@5" ], [ "Italian to English", "@10" ], [ "English to Italian", "@1" ], [ "English to Italian", "@5" ], [ "English to Italian", "@10" ], [ "Italian to English", "@1" ], [ "Italian to English", "@5" ], [ "Italian to English", "@10" ] ], "contents": [ [ "33.8", "48.3", "53.9", "24.9", "41", "47.4", "1", "2.8", "3.9", "2.5", "6.4", "9.1" ], [ "36.1", "52.7", "58.1", "31", "49.9", "57", "29.1", "46.4", "53", "27", "47", "52.3" ], [ "30.9", "48.1", "52.7", "27.7", "45.5", "51", "26.5", "42.5", "48.1", "22.8", "40.1", "45.5" ], [ "36.9", "52.7", "57.9", "32.2", "49.6", "55.7", "27.1", "43.4", "49.3", "26.2", "42.1", "49" ], [ "39.3", "55.3", "60.1", "34.7", "53.5", "59.4", "34.7", "52.6", "58.3", "33.7", "53.3", "59.2" ] ], "metrics_loc": "column", "metrics_type": "['@1', '@5', '@10', '@1', '@5', '@10', '@1', '@5', '@10', '@1', '@5', '@10']", "target_entity": "['IBFA (Ours)']", "table_html": "
English to Italian || @1 English to Italian || @5 English to Italian || @10 Italian to English || @1 Italian to English || @5 Italian to English || @10 English to Italian || @1 English to Italian || @5 English to Italian || @10 Italian to English || @1 Italian to English || @5 Italian to English || @10
Methods || Mikolov et. al. 33.8 48.3 53.9 24.9 41 47.4 1 2.8 3.9 2.5 6.4 9.1
Methods || CCA (Sklearn) 36.1 52.7 58.1 31 49.9 57 29.1 46.4 53 27 47 52.3
Methods || CCA 30.9 48.1 52.7 27.7 45.5 51 26.5 42.5 48.1 22.8 40.1 45.5
Methods || SVD 36.9 52.7 57.9 32.2 49.6 55.7 27.1 43.4 49.3 26.2 42.1 49
Methods || IBFA (Ours) 39.3 55.3 60.1 34.7 53.5 59.4 34.7 52.6 58.3 33.7 53.3 59.2
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P19-1170", "page_no": 6, "dir": "acl2019", "description": "4.1.1 Detailed Comparison to Basic SVD. We present a more detailed comparison to the SVD method described in (Smith et al., 2017). We focus on methods in their base form, that is without post-processing techniques, i.e. cross-domain similarity local scaling (CSLS) (Lample et al., 2018) or inverted softmax (ISF) (Smithet al., 2017). Note that (Smith et al., 2017) used the scikit-learn 2 implementation of CCA, which uses an iterative estimation of partial least squares. This does not give the same results as the standard CCA procedure. In Table 2 we reproduce the results from (Smith et al., 2017) using the dictionaries and embeddings provided by (Dinu et al., 2014)3 and we compare our method (IBFA) using both the expert dictionaries from (Dinu et al., 2014) and the pseudo-dictionaries as constructed in (Smith et al., 2017). We significantly outperform both SVD and CCA, especially when using the pseudo-dictionaries.", "class_sentence": "[2, 2, 2, 2, 2, 1, 1]", "sentences": "['4.1.1 Detailed Comparison to Basic SVD.', 'We present a more detailed comparison to the SVD method described in (Smith et al., 2017).', 'We focus on methods in their base form, that is without post-processing techniques, i.e. cross-domain similarity local scaling (CSLS) (Lample et al., 2018) or inverted softmax (ISF) (Smithet al., 2017).', 'Note that (Smith et al., 2017) used the scikit-learn 2 implementation of CCA, which uses an iterative estimation of partial least squares.', 'This does not give the same results as the standard CCA procedure.', 'In Table 2 we reproduce the results from (Smith et al., 2017) using the dictionaries and embeddings provided by (Dinu et al., 2014)3 and we compare our method (IBFA) using both the expert dictionaries from (Dinu et al., 2014) and the pseudo-dictionaries as constructed in (Smith et al., 2017).', 'We significantly outperform both SVD and CCA, especially when using the pseudo-dictionaries.']", "header_mention": "[None, ['SVD'], None, ['CCA'], ['CCA'], ['IBFA (Ours)'], ['IBFA (Ours)', 'SVD', 'CCA']]", "valid": 1, "image_id": "P19-1170-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparisons without post-processing of methods. Results reproduced from (Smith et al., 2017) for fair comparison. Left: Comparisons using the same expert dictionary as (Smith et al., 2017). Right: Comparisons using the pseudo-dictionary from (Smith et al., 2017).

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n English to Italian || @1English to Italian || @5English to Italian || @10Italian to English || @1Italian to English || @5Italian to English || @10English to Italian || @1English to Italian || @5English to Italian || @10Italian to English || @1Italian to English || @5Italian to English || @10
Methods || Mikolov et. al.33.848.353.924.94147.412.83.92.56.49.1
Methods || CCA (Sklearn)36.152.758.13149.95729.146.453274752.3
Methods || CCA30.948.152.727.745.55126.542.548.122.840.145.5
Methods || SVD36.952.757.932.249.655.727.143.449.326.242.149
Methods || IBFA (Ours)39.355.360.134.753.559.434.752.658.333.753.359.2
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparisons without post-processing of methods. Results reproduced from (Smith et al., 2017) for fair comparison. Left: Comparisons using the same expert dictionary as (Smith et al., 2017). Right: Comparisons using the pseudo-dictionary from (Smith et al., 2017).}\n\\begin{tabular}{lrrrrrrrrrrrr}\n\\toprule\n & English to Italian $\\parallel$ @1 & English to Italian $\\parallel$ @5 & English to Italian $\\parallel$ @10 & Italian to English $\\parallel$ @1 & Italian to English $\\parallel$ @5 & Italian to English $\\parallel$ @10 & English to Italian $\\parallel$ @1.1 & English to Italian $\\parallel$ @5.1 & English to Italian $\\parallel$ @10.1 & Italian to English $\\parallel$ @1.1 & Italian to English $\\parallel$ @5.1 & Italian to English $\\parallel$ @10.1 \\\\\n\\midrule\nMethods $\\parallel$ Mikolov et. al. & 33.8 & 48.3 & 53.9 & 24.9 & 41 & 47.4 & 1 & 2.8 & 3.9 & 2.5 & 6.4 & 9.1 \\\\\nMethods $\\parallel$ CCA (Sklearn) & 36.1 & 52.7 & 58.1 & 31 & 49.9 & 57 & 29.1 & 46.4 & 53 & 27 & 47 & 52.3 \\\\\nMethods $\\parallel$ CCA & 30.9 & 48.1 & 52.7 & 27.7 & 45.5 & 51 & 26.5 & 42.5 & 48.1 & 22.8 & 40.1 & 45.5 \\\\\nMethods $\\parallel$ SVD & 36.9 & 52.7 & 57.9 & 32.2 & 49.6 & 55.7 & 27.1 & 43.4 & 49.3 & 26.2 & 42.1 & 49 \\\\\nMethods $\\parallel$ IBFA (Ours) & 39.3 & 55.3 & 60.1 & 34.7 & 53.5 & 59.4 & 34.7 & 52.6 & 58.3 & 33.7 & 53.3 & 59.2 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Comparisons without post-processing of methods. Results reproduced from (Smith et al., 2017) for fair comparison. Left: Comparisons using the same expert dictionary as (Smith et al., 2017). Right: Comparisons using the pseudo-dictionary from (Smith et al., 2017).\n
\n \n English to Italian || @1\n \n English to Italian || @5\n \n English to Italian || @10\n \n Italian to English || @1\n \n Italian to English || @5\n \n Italian to English || @10\n \n English to Italian || @1\n \n English to Italian || @5\n \n English to Italian || @10\n \n Italian to English || @1\n \n Italian to English || @5\n \n Italian to English || @10\n
\n Methods || Mikolov et. al.\n \n 33.8\n \n 48.3\n \n 53.9\n \n 24.9\n \n 41\n \n 47.4\n \n 1\n \n 2.8\n \n 3.9\n \n 2.5\n \n 6.4\n \n 9.1\n
\n Methods || CCA (Sklearn)\n \n 36.1\n \n 52.7\n \n 58.1\n \n 31\n \n 49.9\n \n 57\n \n 29.1\n \n 46.4\n \n 53\n \n 27\n \n 47\n \n 52.3\n
\n Methods || CCA\n \n 30.9\n \n 48.1\n \n 52.7\n \n 27.7\n \n 45.5\n \n 51\n \n 26.5\n \n 42.5\n \n 48.1\n \n 22.8\n \n 40.1\n \n 45.5\n
\n Methods || SVD\n \n 36.9\n \n 52.7\n \n 57.9\n \n 32.2\n \n 49.6\n \n 55.7\n \n 27.1\n \n 43.4\n \n 49.3\n \n 26.2\n \n 42.1\n \n 49\n
\n Methods || IBFA (Ours)\n \n 39.3\n \n 55.3\n \n 60.1\n \n 34.7\n \n 53.5\n \n 59.4\n \n 34.7\n \n 52.6\n \n 58.3\n \n 33.7\n \n 53.3\n \n 59.2\n
\n" }, { "instance_id": "e548b43a-ea06-41f5-a426-85ba27cdf9ec", "table_id_paper": "P19-1193table_4", "caption": "Human evaluations of ablation study.", "row_header_level": 2, "row_headers": [ [ "Methods", "Full model" ], [ "Methods", "w/o Adversarial" ], [ "Methods", "w/o Memory" ], [ "Methods", "w/o Dynamic" ] ], "column_header_level": 1, "column_headers": [ [ "Consistency" ], [ "Novelty" ], [ "Diversity" ], [ "Coherence" ] ], "contents": [ [ "3.84", "3.24", "3.16", "3.61" ], [ "3.31", "3.07", "3.14", "3.43" ], [ "3.53", "2.73", "2.77", "3.19" ], [ "3.62", "2.91", "2.95", "3.37" ] ], "metrics_loc": "column", "metrics_type": "['Consistency', 'Novelty', 'Diversity', 'Coherence']", "target_entity": "['Methods']", "table_html": "
Consistency Novelty Diversity Coherence
Methods || Full model 3.84 3.24 3.16 3.61
Methods || w/o Adversarial 3.31 3.07 3.14 3.43
Methods || w/o Memory 3.53 2.73 2.77 3.19
Methods || w/o Dynamic 3.62 2.91 2.95 3.37
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P19-1193", "page_no": 7, "dir": "acl2019", "description": "Adversarial training . Another conclusion is that adversarial training can better benefit the model to enhance the topic-consistency of the generated essay compared to memory mechanism. In detail, Table 4 shows that the consistency score given by humans for ablated versions without adversarial training and memory mechanism decline 0.53 and 0.31, respectively. The reason is that the discriminative signal in training not only evaluates the quality of the generated text, but also models its degree of association with the input topics, thus enhancing the topic-consistency.", "class_sentence": "[2, 2, 1, 2]", "sentences": "['Adversarial training .', 'Another conclusion is that adversarial training can better benefit the model to enhance the topic-consistency of the generated essay compared to memory mechanism.', 'In detail, Table 4 shows that the consistency score given by humans for ablated versions without adversarial training and memory mechanism decline 0.53 and 0.31, respectively.', 'The reason is that the discriminative signal in training not only evaluates the quality of the generated text, but also models its degree of association with the input topics, thus enhancing the topic-consistency.']", "header_mention": "[None, None, ['w/o Adversarial', 'w/o Memory', 'Consistency'], None]", "valid": 1, "image_id": "P19-1193-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Human evaluations of ablation study.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n ConsistencyNoveltyDiversityCoherence
Methods || Full model3.843.243.163.61
Methods || w/o Adversarial3.313.073.143.43
Methods || w/o Memory3.532.732.773.19
Methods || w/o Dynamic3.622.912.953.37
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Human evaluations of ablation study.}\n\\begin{tabular}{lrrrr}\n\\toprule\n & Consistency & Novelty & Diversity & Coherence \\\\\n\\midrule\nMethods $\\parallel$ Full model & 3.84 & 3.24 & 3.16 & 3.61 \\\\\nMethods $\\parallel$ w/o Adversarial & 3.31 & 3.07 & 3.14 & 3.43 \\\\\nMethods $\\parallel$ w/o Memory & 3.53 & 2.73 & 2.77 & 3.19 \\\\\nMethods $\\parallel$ w/o Dynamic & 3.62 & 2.91 & 2.95 & 3.37 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Human evaluations of ablation study.\n
\n \n Consistency\n \n Novelty\n \n Diversity\n \n Coherence\n
\n Methods || Full model\n \n 3.84\n \n 3.24\n \n 3.16\n \n 3.61\n
\n Methods || w/o Adversarial\n \n 3.31\n \n 3.07\n \n 3.14\n \n 3.43\n
\n Methods || w/o Memory\n \n 3.53\n \n 2.73\n \n 2.77\n \n 3.19\n
\n Methods || w/o Dynamic\n \n 3.62\n \n 2.91\n \n 2.95\n \n 3.37\n
\n" }, { "instance_id": "83251ea4-1925-4c13-9616-8beff514513a", "table_id_paper": "P19-1230table_2", "caption": "English dev set performance of joint span HPSG parsing. The converted means the corresponding dependency parsing results are from the corresponding constituent parse tree using head rules.", "row_header_level": 2, "row_headers": [ [ "Model", "separate constituent" ], [ "Model", "converted dependency" ], [ "Model", "HPSG Parser joint span lambda = 1.0" ], [ "Model", "HPSG Parser joint span lambda = 0.0" ], [ "Model", "HPSG Parser joint span lambda = 0.5" ], [ "Model", "HPSG Parser converted dependency" ] ], "column_header_level": 1, "column_headers": [ [ "F1" ], [ "UAS" ], [ "LAS" ] ], "contents": [ [ "93.47", "-", "-" ], [ "93.47", "95.06", "93.81" ], [ "93.67", "-", "-" ], [ "-", "95.82", "94.43" ], [ "93.78", "95.92", "94.49" ], [ "93.78", "95.69", "94.45" ] ], "metrics_loc": "column", "metrics_type": "['F1', 'UAS', 'LAS']", "target_entity": "['HPSG Parser joint span lambda = 1.0', 'HPSG Parser joint span lambda = 0.0', 'HPSG Parser joint span lambda = 0.5', 'HPSG Parser converted dependency']", "table_html": "
F1 UAS LAS
Model || separate constituent 93.47 - -
Model || converted dependency 93.47 95.06 93.81
Model || HPSG Parser joint span lambda = 1.0 93.67 - -
Model || HPSG Parser joint span lambda = 0.0 - 95.82 94.43
Model || HPSG Parser joint span lambda = 0.5 93.78 95.92 94.49
Model || HPSG Parser converted dependency 93.78 95.69 94.45
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P19-1230", "page_no": 7, "dir": "acl2019", "description": "When lambda is set to 0 and 1, our joint span HPSG parser works as the dependency-only parser and constituent-only parser respectively. Table 2 shows that even in such a work mode, our HPSG parser still outperforms the separate constituent parser in terms of either constituent and dependency parsing performance. As lambda is set to 0.5, our HPSG parser will give constituent and dependency structures at the same time, which are shown better than the work alone mode of either constituent or dependency parsing. Besides, the comparison also shows that the directly predicted dependencies from our model are slightly better than those converted from the predicted constituent parse trees.", "class_sentence": "[2, 1, 1, 1]", "sentences": "['When lambda is set to 0 and 1, our joint span HPSG parser works as the dependency-only parser and constituent-only parser respectively.', 'Table 2 shows that even in such a work mode, our HPSG parser still outperforms the separate constituent parser in terms of either constituent and dependency parsing performance.', 'As lambda is set to 0.5, our HPSG parser will give constituent and dependency structures at the same time, which are shown better than the work alone mode of either constituent or dependency parsing.', 'Besides, the comparison also shows that the directly predicted dependencies from our model are slightly better than those converted from the predicted constituent parse trees.']", "header_mention": "[['HPSG Parser joint span lambda = 0.0', 'HPSG Parser joint span lambda = 1.0'], ['HPSG Parser joint span lambda = 0.0', 'HPSG Parser joint span lambda = 1.0'], ['HPSG Parser joint span lambda = 0.5', 'HPSG Parser joint span lambda = 0.0', 'HPSG Parser joint span lambda = 1.0'], ['HPSG Parser converted dependency', 'converted dependency']]", "valid": 1, "image_id": "P19-1230-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

English dev set performance of joint span HPSG parsing. The converted means the corresponding dependency parsing results are from the corresponding constituent parse tree using head rules.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n F1UASLAS
Model || separate constituent93.47--
Model || converted dependency93.4795.0693.81
Model || HPSG Parser joint span lambda = 1.093.67--
Model || HPSG Parser joint span lambda = 0.0-95.8294.43
Model || HPSG Parser joint span lambda = 0.593.7895.9294.49
Model || HPSG Parser converted dependency93.7895.6994.45
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{English dev set performance of joint span HPSG parsing. The converted means the corresponding dependency parsing results are from the corresponding constituent parse tree using head rules.}\n\\begin{tabular}{llll}\n\\toprule\n & F1 & UAS & LAS \\\\\n\\midrule\nModel $\\parallel$ separate constituent & 93.47 & - & - \\\\\nModel $\\parallel$ converted dependency & 93.47 & 95.06 & 93.81 \\\\\nModel $\\parallel$ HPSG Parser joint span lambda $=$ 1.0 & 93.67 & - & - \\\\\nModel $\\parallel$ HPSG Parser joint span lambda $=$ 0.0 & - & 95.82 & 94.43 \\\\\nModel $\\parallel$ HPSG Parser joint span lambda $=$ 0.5 & 93.78 & 95.92 & 94.49 \\\\\nModel $\\parallel$ HPSG Parser converted dependency & 93.78 & 95.69 & 94.45 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: English dev set performance of joint span HPSG parsing. The converted means the corresponding dependency parsing results are from the corresponding constituent parse tree using head rules.\n
\n \n F1\n \n UAS\n \n LAS\n
\n Model || separate constituent\n \n 93.47\n \n -\n \n -\n
\n Model || converted dependency\n \n 93.47\n \n 95.06\n \n 93.81\n
\n Model || HPSG Parser joint span lambda = 1.0\n \n 93.67\n \n -\n \n -\n
\n Model || HPSG Parser joint span lambda = 0.0\n \n -\n \n 95.82\n \n 94.43\n
\n Model || HPSG Parser joint span lambda = 0.5\n \n 93.78\n \n 95.92\n \n 94.49\n
\n Model || HPSG Parser converted dependency\n \n 93.78\n \n 95.69\n \n 94.45\n
\n" }, { "instance_id": "2c20ec29-a4a9-4df5-843c-527dd5e2a7bd", "table_id_paper": "P19-1276table_5", "caption": "Averaged slot coherence results.", "row_header_level": 2, "row_headers": [ [ "Method", "Nguyen et al. (2015)" ], [ "Method", "ODEE-F" ], [ "Method", "ODEE-FE" ], [ "Method", "ODEE-FER" ] ], "column_header_level": 1, "column_headers": [ [ "Ave Slot Coherence" ] ], "contents": [ [ "0.1" ], [ "0.1" ], [ "0.16" ], [ "0.18" ] ], "metrics_loc": "column", "metrics_type": "['Ave Slot Coherence']", "target_entity": "['ODEE-F', 'ODEE-FE', 'ODEE-FER']", "table_html": "
Ave Slot Coherence
Method || Nguyen et al. (2015) 0.1
Method || ODEE-F 0.1
Method || ODEE-FE 0.16
Method || ODEE-FER 0.18
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "P19-1276", "page_no": 7, "dir": "acl2019", "description": "Slot Coherence. Table 5 shows the comparison of averaged slot coherence results over all the slots in the schemas. . Note that we do not report the slot coherence for the Clusteringmodel because it does not output the top-N head words in eachslot. . The averaged slot coherence of ODEE-FER is the highest, which is consistent with the conclusion from Table 4. . The averaged slot coherence of ODEE-F is comparable to that of Nguyen et al.(2015)(p= 0.3415), which again demonstrates that the contextual features are a strong alterna-tive to discrete features. The scores of ODEE-FE(p= 0.06) and ODEE-FER(p= 10^5) are both higher than that of ODEE-F, which proves that the latent event type is critical in ODEE.", "class_sentence": "[2, 1, 2, 1, 1, 1]", "sentences": "['Slot Coherence.', 'Table 5 shows the comparison of averaged slot coherence results over all the slots in the schemas. .', 'Note that we do not report the slot coherence for the Clusteringmodel because it does not output the top-N head words in eachslot. .', 'The averaged slot coherence of ODEE-FER is the highest, which is consistent with the conclusion from Table 4. .', 'The averaged slot coherence of ODEE-F is comparable to that of Nguyen et al.(2015)(p= 0.3415), which again demonstrates that the contextual features are a strong alterna-tive to discrete features.', 'The scores of ODEE-FE(p= 0.06) and ODEE-FER(p= 10^5) are both higher than that of ODEE-F, which proves that the latent event type is critical in ODEE.']", "header_mention": "[None, ['Ave Slot Coherence'], None, ['ODEE-FER'], ['Ave Slot Coherence', 'ODEE-F', 'Nguyen et al. (2015)'], ['ODEE-FE', 'ODEE-FER', 'ODEE-F']]", "valid": 1, "image_id": "P19-1276-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Averaged slot coherence results.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Ave Slot Coherence
Method || Nguyen et al. (2015)0.1
Method || ODEE-F0.1
Method || ODEE-FE0.16
Method || ODEE-FER0.18
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Averaged slot coherence results.}\n\\begin{tabular}{lr}\n\\toprule\n & Ave Slot Coherence \\\\\n\\midrule\nMethod $\\parallel$ Nguyen et al. (2015) & 0.1 \\\\\nMethod $\\parallel$ ODEE-F & 0.1 \\\\\nMethod $\\parallel$ ODEE-FE & 0.16 \\\\\nMethod $\\parallel$ ODEE-FER & 0.18 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Averaged slot coherence results.\n
\n \n Ave Slot Coherence\n
\n Method || Nguyen et al. (2015)\n \n 0.1\n
\n Method || ODEE-F\n \n 0.1\n
\n Method || ODEE-FE\n \n 0.16\n
\n Method || ODEE-FER\n \n 0.18\n
\n" }, { "instance_id": "eb6c0059-3a5a-40c3-9a3a-b77efa0e828e", "table_id_paper": "P19-1290table_1", "caption": "Performance of various models on EN-DE and EN-FR translation tasks.", "row_header_level": 2, "row_headers": [ [ "Architecture", "Vaswani et al. (2017) Transformer big" ], [ "Architecture", "Wu et al. (2018) Transformer big + sequence-loss" ], [ "Architecture", "Yang et al. (2018) Transformer big + localness" ], [ "Architecture", "this work Transformer big + hard-attention" ] ], "column_header_level": 2, "column_headers": [ [ "BLEU", "EN-DE" ], [ "BLEU", "EN-FR" ] ], "contents": [ [ "28.4", "41" ], [ "28.75", "41.47" ], [ "28.89", "n/a" ], [ "29.29", "42.26" ] ], "metrics_loc": "column", "metrics_type": "['BLEU', 'BLEU']", "target_entity": "['this work Transformer big + hard-attention']", "table_html": "
BLEU || EN-DE BLEU || EN-FR
Architecture || Vaswani et al. (2017) Transformer big 28.4 41
Architecture || Wu et al. (2018) Transformer big + sequence-loss 28.75 41.47
Architecture || Yang et al. (2018) Transformer big + localness 28.89 n/a
Architecture || this work Transformer big + hard-attention 29.29 42.26
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P19-1290", "page_no": 4, "dir": "acl2019", "description": "Main Results . Table 1 shows the performance of various models on EN-DE and EN-FR translation tasks. These test set case sensitive BLEU scores are obtained using SacreBLEU toolkit1 (Post, 2018). The BLEU score difference between our hard-attention based Transformer model and the original soft-attention based Transformer model indicates the effectiveness of selecting a few relevant source tokens for each target token. The performance gap between our method and sequence loss based Transformer (Wu et al., 2018) shows that the improvements are indeed coming from the hard-attention mechanism. Our approach of incorporating hard-attention into decoder\u2019s top selfattention layer to select relevant tokens yielded better results compared to the Localness SelfAttention (Yang et al., 2018) approach of incorporating localness bias only to lower self-attention layers. It can be noted that our model achieved 29.29 and 42.26 BLEU points on EN-DE and ENFR tasks respectively \u2013 surpassing the previously published models.", "class_sentence": "[2, 1, 2, 1, 1, 1, 1]", "sentences": "['Main Results .', 'Table 1 shows the performance of various models on EN-DE and EN-FR translation tasks.', 'These test set case sensitive BLEU scores are obtained using SacreBLEU toolkit1 (Post, 2018).', 'The BLEU score difference between our hard-attention based Transformer model and the original soft-attention based Transformer model indicates the effectiveness of selecting a few relevant source tokens for each target token.', 'The performance gap between our method and sequence loss based Transformer (Wu et al., 2018) shows that the improvements are indeed coming from the hard-attention mechanism.', 'Our approach of incorporating hard-attention into decoder\u2019s top selfattention layer to select relevant tokens yielded better results compared to the Localness SelfAttention (Yang et al., 2018) approach of incorporating localness bias only to lower self-attention layers.', 'It can be noted that our model achieved 29.29 and 42.26 BLEU points on EN-DE and ENFR tasks respectively \u2013 surpassing the previously published models.']", "header_mention": "[None, ['EN-DE', 'EN-FR'], None, ['this work Transformer big + hard-attention'], ['this work Transformer big + hard-attention', 'Wu et al. (2018) Transformer big + sequence-loss'], ['this work Transformer big + hard-attention', 'Yang et al. (2018) Transformer big + localness'], ['this work Transformer big + hard-attention', 'EN-DE', 'EN-FR']]", "valid": 1, "image_id": "P19-1290-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance of various models on EN-DE and EN-FR translation tasks.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n BLEU || EN-DEBLEU || EN-FR
Architecture || Vaswani et al. (2017) Transformer big28.441
Architecture || Wu et al. (2018) Transformer big + sequence-loss28.7541.47
Architecture || Yang et al. (2018) Transformer big + localness28.89n/a
Architecture || this work Transformer big + hard-attention29.2942.26
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance of various models on EN-DE and EN-FR translation tasks.}\n\\begin{tabular}{lrl}\n\\toprule\n & BLEU $\\parallel$ EN-DE & BLEU $\\parallel$ EN-FR \\\\\n\\midrule\nArchitecture $\\parallel$ Vaswani et al. (2017) Transformer big & 28.4 & 41 \\\\\nArchitecture $\\parallel$ Wu et al. (2018) Transformer big + sequence-loss & 28.75 & 41.47 \\\\\nArchitecture $\\parallel$ Yang et al. (2018) Transformer big + localness & 28.89 & \\\\\nArchitecture $\\parallel$ this work Transformer big + hard-attention & 29.29 & 42.26 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Performance of various models on EN-DE and EN-FR translation tasks.\n
\n \n BLEU || EN-DE\n \n BLEU || EN-FR\n
\n Architecture || Vaswani et al. (2017) Transformer big\n \n 28.4\n \n 41\n
\n Architecture || Wu et al. (2018) Transformer big + sequence-loss\n \n 28.75\n \n 41.47\n
\n Architecture || Yang et al. (2018) Transformer big + localness\n \n 28.89\n \n n/a\n
\n Architecture || this work Transformer big + hard-attention\n \n 29.29\n \n 42.26\n
\n" }, { "instance_id": "3a67b7cb-cfb3-4774-9c50-577f4a640e77", "table_id_paper": "P19-1305table_5", "caption": "Validation set performances of using different layers for attention relay.", "row_header_level": 2, "row_headers": [ [ "Layer", "1" ], [ "Layer", "2" ], [ "Layer", "3" ], [ "Layer", "4" ], [ "Layer", "5" ], [ "Layer", "6" ] ], "column_header_level": 1, "column_headers": [ [ "ROUGE-1" ], [ "ROUGE-2" ], [ "ROUGE-L" ] ], "contents": [ [ "44.7", "21.8", "41.6" ], [ "44.7", "22.3", "41.7" ], [ "45", "22", "41.8" ], [ "44.9", "22.1", "41.3" ], [ "44.9", "22.1", "41.9" ], [ "45.1", "22.3", "42" ] ], "metrics_loc": "column", "metrics_type": "['ROUGE-1', 'ROUGE-2', 'ROUGE-L']", "target_entity": "['Layer']", "table_html": "
ROUGE-1 ROUGE-2 ROUGE-L
Layer || 1 44.7 21.8 41.6
Layer || 2 44.7 22.3 41.7
Layer || 3 45 22 41.8
Layer || 4 44.9 22.1 41.3
Layer || 5 44.9 22.1 41.9
Layer || 6 45.1 22.3 42
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "P19-1305", "page_no": 8, "dir": "acl2019", "description": "Layers for Attention Relay. Transformer architecture used in our experiment is with six layers on both encoder and decoder. Attention relay can take place on each layer. Since each layer has eight heads for attention computation, we average the weights of all eight heads in the same layer. We study the attention relay effects on all six layers. The results in Table 5 show that relaying attention on the last layer achieves the best performance.", "class_sentence": "[2, 2, 2, 2, 1, 1]", "sentences": "['Layers for Attention Relay.', 'Transformer architecture used in our experiment is with six layers on both encoder and decoder.', 'Attention relay can take place on each layer.', 'Since each layer has eight heads for attention computation, we average the weights of all eight heads in the same layer.', 'We study the attention relay effects on all six layers.', 'The results in Table 5 show that relaying attention on the last layer achieves the best performance.']", "header_mention": "[None, ['Layer'], ['Layer'], ['Layer'], ['Layer'], ['6', 'Layer']]", "valid": 1, "image_id": "P19-1305-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Validation set performances of using different layers for attention relay.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n ROUGE-1ROUGE-2ROUGE-L
Layer || 144.721.841.6
Layer || 244.722.341.7
Layer || 3452241.8
Layer || 444.922.141.3
Layer || 544.922.141.9
Layer || 645.122.342
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Validation set performances of using different layers for attention relay.}\n\\begin{tabular}{lrrr}\n\\toprule\n & ROUGE-1 & ROUGE-2 & ROUGE-L \\\\\n\\midrule\nLayer $\\parallel$ 1 & 44.7 & 21.8 & 41.6 \\\\\nLayer $\\parallel$ 2 & 44.7 & 22.3 & 41.7 \\\\\nLayer $\\parallel$ 3 & 45 & 22 & 41.8 \\\\\nLayer $\\parallel$ 4 & 44.9 & 22.1 & 41.3 \\\\\nLayer $\\parallel$ 5 & 44.9 & 22.1 & 41.9 \\\\\nLayer $\\parallel$ 6 & 45.1 & 22.3 & 42 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Validation set performances of using different layers for attention relay.\n
\n \n ROUGE-1\n \n ROUGE-2\n \n ROUGE-L\n
\n Layer || 1\n \n 44.7\n \n 21.8\n \n 41.6\n
\n Layer || 2\n \n 44.7\n \n 22.3\n \n 41.7\n
\n Layer || 3\n \n 45\n \n 22\n \n 41.8\n
\n Layer || 4\n \n 44.9\n \n 22.1\n \n 41.3\n
\n Layer || 5\n \n 44.9\n \n 22.1\n \n 41.9\n
\n Layer || 6\n \n 45.1\n \n 22.3\n \n 42\n
\n" }, { "instance_id": "68e680a8-d754-4897-89fa-0bd65942a641", "table_id_paper": "P19-1309table_3", "caption": "BUCC results (F1) on the test set. We use the ratio function with maximum score retrieval and the filtering threshold optimized on the training set.", "row_header_level": 1, "row_headers": [ [ "Azpeitia et al. (2017)" ], [ "Azpeitia et al. (2018)" ], [ "Bouamor and Sajjad (2018)" ], [ "Schwenk (2018)" ], [ "Proposed method (Europarl)" ], [ "Proposed method (UN)" ] ], "column_header_level": 1, "column_headers": [ [ "en-de" ], [ "en-fr" ], [ "en-ru" ], [ "en-zh" ] ], "contents": [ [ "83.7", "79.5", "-", "-" ], [ "85.5", "81.5", "81.3", "77.5" ], [ "-", "76", "-", "-" ], [ "76.9", "75.8", "73.8", "71.6" ], [ "95.6", "92.9", "-", "-" ], [ "-", "-", "92", "92.6" ] ], "metrics_loc": "column", "metrics_type": "['F1', 'F1', 'F1', 'F1']", "target_entity": "['Proposed method (Europarl)', 'Proposed method (UN)']", "table_html": "
en-de en-fr en-ru en-zh
Azpeitia et al. (2017) 83.7 79.5 - -
Azpeitia et al. (2018) 85.5 81.5 81.3 77.5
Bouamor and Sajjad (2018) - 76 - -
Schwenk (2018) 76.9 75.8 73.8 71.6
Proposed method (Europarl) 95.6 92.9 - -
Proposed method (UN) - - 92 92.6
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "P19-1309", "page_no": 4, "dir": "acl2019", "description": "Table 3 reports the results on the test set for both the Europarl and the UN model in comparison to previous work.9 . Our proposed system outperforms all previous methods by a large margin, obtaining improvements of 10-15 F1 points and showing very consistent performance across different languages, including distant ones.", "class_sentence": "[1, 1]", "sentences": "['Table 3 reports the results on the test set for both the Europarl and the UN model in comparison to previous work.9 .', 'Our proposed system outperforms all previous methods by a large margin, obtaining improvements of 10-15 F1 points and showing very consistent performance across different languages, including distant ones.']", "header_mention": "[None, ['Proposed method (Europarl)', 'Proposed method (UN)']]", "valid": 1, "image_id": "P19-1309-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

BUCC results (F1) on the test set. We use the ratio function with maximum score retrieval and the filtering threshold optimized on the training set.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n en-deen-fren-ruen-zh
Azpeitia et al. (2017)83.779.5--
Azpeitia et al. (2018)85.581.581.377.5
Bouamor and Sajjad (2018)-76--
Schwenk (2018)76.975.873.871.6
Proposed method (Europarl)95.692.9--
Proposed method (UN)--9292.6
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{BUCC results (F1) on the test set. We use the ratio function with maximum score retrieval and the filtering threshold optimized on the training set.}\n\\begin{tabular}{lllll}\n\\toprule\n & en-de & en-fr & en-ru & en-zh \\\\\n\\midrule\nAzpeitia et al. (2017) & 83.7 & 79.5 & - & - \\\\\nAzpeitia et al. (2018) & 85.5 & 81.5 & 81.3 & 77.5 \\\\\nBouamor and Sajjad (2018) & - & 76 & - & - \\\\\nSchwenk (2018) & 76.9 & 75.8 & 73.8 & 71.6 \\\\\nProposed method (Europarl) & 95.6 & 92.9 & - & - \\\\\nProposed method (UN) & - & - & 92 & 92.6 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: BUCC results (F1) on the test set. We use the ratio function with maximum score retrieval and the filtering threshold optimized on the training set.\n
\n \n en-de\n \n en-fr\n \n en-ru\n \n en-zh\n
\n Azpeitia et al. (2017)\n \n 83.7\n \n 79.5\n \n -\n \n -\n
\n Azpeitia et al. (2018)\n \n 85.5\n \n 81.5\n \n 81.3\n \n 77.5\n
\n Bouamor and Sajjad (2018)\n \n -\n \n 76\n \n -\n \n -\n
\n Schwenk (2018)\n \n 76.9\n \n 75.8\n \n 73.8\n \n 71.6\n
\n Proposed method (Europarl)\n \n 95.6\n \n 92.9\n \n -\n \n -\n
\n Proposed method (UN)\n \n -\n \n -\n \n 92\n \n 92.6\n
\n" }, { "instance_id": "7b35fdbf-4050-4ddb-af7b-e2fbfefc09f2", "table_id_paper": "P19-1316table_1", "caption": "Baseline (Cordeiro et al., 2016) results on the reduced version of three gold-standard datasets ordered in decreasing overall performance along with the results of using only Poincar\u00b4e embedding.", "row_header_level": 2, "row_headers": [ [ "Base. Model", "W2V-CBOW" ], [ "Base. Model", "W2V-SG" ], [ "Base. Model", "GloVe" ], [ "Base. Model", "PPMI-SVD" ], [ "Base. Model", "Poincare" ] ], "column_header_level": 1, "column_headers": [ [ "RD-R" ], [ "RD++-R" ], [ "FD-R" ] ], "contents": [ [ "0.8045", "0.6964", "0.3405" ], [ "0.8034", "0.6963", "0.3396" ], [ "0.7604", "0.6487", "0.262" ], [ "0.7484", "0.6468", "0.2428" ], [ "0.6023", "0.4765", "0.2007" ] ], "metrics_loc": "column", "metrics_type": "['correlation', 'correlation', 'correlation']", "target_entity": "['Poincare', 'W2V-CBOW']", "table_html": "
RD-R RD++-R FD-R
Base. Model || W2V-CBOW 0.8045 0.6964 0.3405
Base. Model || W2V-SG 0.8034 0.6963 0.3396
Base. Model || GloVe 0.7604 0.6487 0.262
Base. Model || PPMI-SVD 0.7484 0.6468 0.2428
Base. Model || Poincare 0.6023 0.4765 0.2007
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P19-1316", "page_no": 6, "dir": "acl2019", "description": "We compare the performance of the baseline models (Cordeiro et al., 2016) and Poincare embeddings as a single signal on the reduced version of the three gold standard datasets: RD-R (79 instances), RD++-R (146 instances), FD-R (780 instances) in order to closely examine the influence of Poincare embeddings. Table 1 shows the performance for all the baselines in terms of Spearman's rank correlation. We observe that W2V-CBOW model produces the best performance across all the three datasets and W2V-SG achieves the second-best performance. As noted in the table, the Poincare embeddings on their own perform worse than all the other baselines. Further, since our final model is based on an interpolation between Poincare embeddings and W2VCBOW, we also attempted interpolation between other four baseline models, but the best results were always close to the better of the two models, and are not reported here.", "class_sentence": "[2, 1, 1, 1, 2]", "sentences": "['We compare the performance of the baseline models (Cordeiro et al., 2016) and Poincare embeddings as a single signal on the reduced version of the three gold standard datasets: RD-R (79 instances), RD++-R (146 instances), FD-R (780 instances) in order to closely examine the influence of Poincare embeddings.', \"Table 1 shows the performance for all the baselines in terms of Spearman's rank correlation.\", 'We observe that W2V-CBOW model produces the best performance across all the three datasets and W2V-SG achieves the second-best performance.', 'As noted in the table, the Poincare embeddings on their own perform worse than all the other baselines.', 'Further, since our final model is based on an interpolation between Poincare embeddings and W2VCBOW, we also attempted interpolation between other four baseline models, but the best results were always close to the better of the two models, and are not reported here.']", "header_mention": "[['Poincare', 'RD-R', 'RD++-R', 'FD-R'], None, ['W2V-CBOW', 'W2V-SG'], ['Poincare'], None]", "valid": 1, "image_id": "P19-1316-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Baseline (Cordeiro et al., 2016) results on the reduced version of three gold-standard datasets ordered in decreasing overall performance along with the results of using only Poincar\u00b4e embedding.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n RD-RRD++-RFD-R
Base. Model || W2V-CBOW0.80450.69640.3405
Base. Model || W2V-SG0.80340.69630.3396
Base. Model || GloVe0.76040.64870.262
Base. Model || PPMI-SVD0.74840.64680.2428
Base. Model || Poincare0.60230.47650.2007
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Baseline (Cordeiro et al., 2016) results on the reduced version of three gold-standard datasets ordered in decreasing overall performance along with the results of using only Poincar\u00b4e embedding.}\n\\begin{tabular}{lrrr}\n\\toprule\n & RD-R & RD++-R & FD-R \\\\\n\\midrule\nBase. Model $\\parallel$ W2V-CBOW & 0.8045 & 0.6964 & 0.3405 \\\\\nBase. Model $\\parallel$ W2V-SG & 0.8034 & 0.6963 & 0.3396 \\\\\nBase. Model $\\parallel$ GloVe & 0.7604 & 0.6487 & 0.262 \\\\\nBase. Model $\\parallel$ PPMI-SVD & 0.7484 & 0.6468 & 0.2428 \\\\\nBase. Model $\\parallel$ Poincare & 0.6023 & 0.4765 & 0.2007 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Baseline (Cordeiro et al., 2016) results on the reduced version of three gold-standard datasets ordered in decreasing overall performance along with the results of using only Poincar\u00b4e embedding.\n
\n \n RD-R\n \n RD++-R\n \n FD-R\n
\n Base. Model || W2V-CBOW\n \n 0.8045\n \n 0.6964\n \n 0.3405\n
\n Base. Model || W2V-SG\n \n 0.8034\n \n 0.6963\n \n 0.3396\n
\n Base. Model || GloVe\n \n 0.7604\n \n 0.6487\n \n 0.262\n
\n Base. Model || PPMI-SVD\n \n 0.7484\n \n 0.6468\n \n 0.2428\n
\n Base. Model || Poincare\n \n 0.6023\n \n 0.4765\n \n 0.2007\n
\n" }, { "instance_id": "a81d11d2-3515-439c-863c-5864495e2a96", "table_id_paper": "P19-1332table_8", "caption": "Human Evaluation in WikiAnswers\u2192Quora", "row_header_level": 2, "row_headers": [ [ "Models", "MTL+Copy" ], [ "Models", "Na\u00efve DNPG" ], [ "Models", "Adapted DNPG" ], [ "Models", "Reference" ] ], "column_header_level": 1, "column_headers": [ [ "Mean Rank" ], [ "Agreement" ] ], "contents": [ [ "3.22", "0.446" ], [ "3.13", "0.323" ], [ "1.79", "0.383" ], [ "1.48", "0.338" ] ], "metrics_loc": "column", "metrics_type": "['Mean Rank', 'Agreement']", "target_entity": "['MTL+Copy', 'Na\u00efve DNPG', 'Adapted DNPG']", "table_html": "
Mean Rank Agreement
Models || MTL+Copy 3.22 0.446
Models || Na\u00efve DNPG 3.13 0.323
Models || Adapted DNPG 1.79 0.383
Models || Reference 1.48 0.338
", "table_name": "Table 8", "table_id": "table_8", "paper_id": "P19-1332", "page_no": 8, "dir": "acl2019", "description": "MTL+Copy, DNPG and Adapted DNPG as well as the reference. The evaluators are asked to rank the candidates from 1 (best) to 4 (worst) by their readability, accuracy and surface dissimilarity to the input sentence. The detailed evaluation guide can be found in Appendix B. Table 8 shows the mean rank and inter-annotator agreement (Cohen's kappa) of each model. Adapted DNPG again significantly outperforms MTL+Copy by a large margin (p-value < 0.01). The performance of the original DNPG and MTL+Copy has no significant difference (p-value = 0.18). All of the interannotator agreement is regarded as fair or above.", "class_sentence": "[1, 2, 0, 1, 1, 1, 1]", "sentences": "['MTL+Copy, DNPG and Adapted DNPG as well as the reference.', 'The evaluators are asked to rank the candidates from 1 (best) to 4 (worst) by their readability, accuracy and surface dissimilarity to the input sentence.', 'The detailed evaluation guide can be found in Appendix B.', \"Table 8 shows the mean rank and inter-annotator agreement (Cohen's kappa) of each model.\", 'Adapted DNPG again significantly outperforms MTL+Copy by a large margin (p-value < 0.01).', 'The performance of the original DNPG and MTL+Copy has no significant difference (p-value = 0.18).', 'All of the interannotator agreement is regarded as fair or above.']", "header_mention": "[['MTL+Copy', 'Na\u00efve DNPG', 'Adapted DNPG'], None, None, ['Mean Rank', 'Agreement'], ['Adapted DNPG', 'MTL+Copy'], ['Na\u00efve DNPG', 'MTL+Copy'], ['Agreement']]", "valid": 1, "image_id": "P19-1332-Table8-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Human Evaluation in WikiAnswers\u2192Quora

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Mean RankAgreement
Models || MTL+Copy3.220.446
Models || Na\u00efve DNPG3.130.323
Models || Adapted DNPG1.790.383
Models || Reference1.480.338
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{7}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Human Evaluation in WikiAnswers$\\rightarrow$Quora}\n\\begin{tabular}{lrr}\n\\toprule\n & Mean Rank & Agreement \\\\\n\\midrule\nModels $\\parallel$ MTL+Copy & 3.22 & 0.446 \\\\\nModels $\\parallel$ Na\u00efve DNPG & 3.13 & 0.323 \\\\\nModels $\\parallel$ Adapted DNPG & 1.79 & 0.383 \\\\\nModels $\\parallel$ Reference & 1.48 & 0.338 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 8: Human Evaluation in WikiAnswers\u2192Quora\n
\n \n Mean Rank\n \n Agreement\n
\n Models || MTL+Copy\n \n 3.22\n \n 0.446\n
\n Models || Na\u00efve DNPG\n \n 3.13\n \n 0.323\n
\n Models || Adapted DNPG\n \n 1.79\n \n 0.383\n
\n Models || Reference\n \n 1.48\n \n 0.338\n
\n" }, { "instance_id": "6d1e89e5-4fa6-4f76-9eed-a5005eba716a", "table_id_paper": "P19-1370table_2", "caption": "Evaluation results of co-teaching initialized with different networks.", "row_header_level": 1, "row_headers": [ [ "SMN-Pre-training" ], [ "SMN-Co-teaching" ], [ "DAM-Pre-training" ], [ "DAM-Co-teaching" ] ], "column_header_level": 2, "column_headers": [ [ "Douban (Margin)", "MAP" ], [ "Douban (Margin)", "MRR" ], [ "Douban (Margin)", "P@1" ], [ "Douban (Margin)", "R10@1" ], [ "Douban (Margin)", "R10@2" ], [ "Douban (Margin)", "R10@5" ], [ "ECD (Curriculum)", "MAP" ], [ "ECD (Curriculum)", "MRR" ], [ "ECD (Curriculum)", "P@1" ], [ "ECD (Curriculum)", "R10@1" ], [ "ECD (Curriculum)", "R10@2" ], [ "ECD (Curriculum)", "R10@5" ] ], "contents": [ [ "0.527", "0.57", "0.396", "0.236", "0.392", "0.734", "0.662", "0.742", "0.598", "0.302", "0.464", "0.757" ], [ "0.558", "0.602", "0.42", "0.255", "0.431", "0.787", "0.674", "0.765", "0.626", "0.322", "0.485", "0.779" ], [ "0.552", "0.605", "0.426", "0.258", "0.408", "0.766", "0.685", "0.756", "0.621", "0.325", "0.491", "0.772" ], [ "0.57", "0.617", "0.438", "0.27", "0.455", "0.781", "0.696", "0.775", "0.652", "0.341", "0.499", "0.784" ] ], "metrics_loc": "column", "metrics_type": "['MAP', 'MRR', 'P@1', 'R10@1', 'R10@2', 'R10@5', 'MAP', 'MRR', 'P@1', 'R10@1', 'R10@2', 'R10@5']", "target_entity": "['SMN-Co-teaching', 'DAM-Co-teaching']", "table_html": "
Douban (Margin) || MAP Douban (Margin) || MRR Douban (Margin) || P@1 Douban (Margin) || R10@1 Douban (Margin) || R10@2 Douban (Margin) || R10@5 ECD (Curriculum) || MAP ECD (Curriculum) || MRR ECD (Curriculum) || P@1 ECD (Curriculum) || R10@1 ECD (Curriculum) || R10@2 ECD (Curriculum) || R10@5
SMN-Pre-training 0.527 0.57 0.396 0.236 0.392 0.734 0.662 0.742 0.598 0.302 0.464 0.757
SMN-Co-teaching 0.558 0.602 0.42 0.255 0.431 0.787 0.674 0.765 0.626 0.322 0.485 0.779
DAM-Pre-training 0.552 0.605 0.426 0.258 0.408 0.766 0.685 0.756 0.621 0.325 0.491 0.772
DAM-Co-teaching 0.57 0.617 0.438 0.27 0.455 0.781 0.696 0.775 0.652 0.341 0.499 0.784
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P19-1370", "page_no": 7, "dir": "acl2019", "description": "Answer to Q2: as a case study of co-teaching with two networks in different capabilities, we initialize model A and model B with DAM and SMN respectively, and select teaching with dynamic margins for Douban and teaching with dynamic data curriculum for ECD (i.e., the best strategies for the two data sets when co-teaching is initialized with one network). Table 2 shows comparison between models before/after co-teaching. We find that co-teaching is still effective when starting from two networks, as both SMN and DAM get improved on the two data sets. Despite the improvement, it is still better to learn the two networks one by one, as co-teaching with two networks cannot bring more improvement than coteaching with one network, and the performance.", "class_sentence": "[2, 1, 1, 2]", "sentences": "['Answer to Q2: as a case study of co-teaching with two networks in different capabilities, we initialize model A and model B with DAM and SMN respectively, and select teaching with dynamic margins for Douban and teaching with dynamic data curriculum for ECD (i.e., the best strategies for the two data sets when co-teaching is initialized with one network).', 'Table 2 shows comparison between models before/after co-teaching.', 'We find that co-teaching is still effective when starting from two networks, as both SMN and DAM get improved on the two data sets.', 'Despite the improvement, it is still better to learn the two networks one by one, as co-teaching with two networks cannot bring more improvement than coteaching with one network, and the performance.']", "header_mention": "[['DAM-Co-teaching', 'SMN-Co-teaching'], None, ['DAM-Co-teaching', 'SMN-Co-teaching'], None]", "valid": 1, "image_id": "P19-1370-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Evaluation results of co-teaching initialized with different networks.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Douban (Margin) || MAPDouban (Margin) || MRRDouban (Margin) || P@1Douban (Margin) || R10@1Douban (Margin) || R10@2Douban (Margin) || R10@5ECD (Curriculum) || MAPECD (Curriculum) || MRRECD (Curriculum) || P@1ECD (Curriculum) || R10@1ECD (Curriculum) || R10@2ECD (Curriculum) || R10@5
SMN-Pre-training0.5270.570.3960.2360.3920.7340.6620.7420.5980.3020.4640.757
SMN-Co-teaching0.5580.6020.420.2550.4310.7870.6740.7650.6260.3220.4850.779
DAM-Pre-training0.5520.6050.4260.2580.4080.7660.6850.7560.6210.3250.4910.772
DAM-Co-teaching0.570.6170.4380.270.4550.7810.6960.7750.6520.3410.4990.784
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Evaluation results of co-teaching initialized with different networks.}\n\\begin{tabular}{lrrrrrrrrrrrr}\n\\toprule\n & Douban (Margin) $\\parallel$ MAP & Douban (Margin) $\\parallel$ MRR & Douban (Margin) $\\parallel$ P@1 & Douban (Margin) $\\parallel$ R10@1 & Douban (Margin) $\\parallel$ R10@2 & Douban (Margin) $\\parallel$ R10@5 & ECD (Curriculum) $\\parallel$ MAP & ECD (Curriculum) $\\parallel$ MRR & ECD (Curriculum) $\\parallel$ P@1 & ECD (Curriculum) $\\parallel$ R10@1 & ECD (Curriculum) $\\parallel$ R10@2 & ECD (Curriculum) $\\parallel$ R10@5 \\\\\n\\midrule\nSMN-Pre-training & 0.527 & 0.57 & 0.396 & 0.236 & 0.392 & 0.734 & 0.662 & 0.742 & 0.598 & 0.302 & 0.464 & 0.757 \\\\\nSMN-Co-teaching & 0.558 & 0.602 & 0.42 & 0.255 & 0.431 & 0.787 & 0.674 & 0.765 & 0.626 & 0.322 & 0.485 & 0.779 \\\\\nDAM-Pre-training & 0.552 & 0.605 & 0.426 & 0.258 & 0.408 & 0.766 & 0.685 & 0.756 & 0.621 & 0.325 & 0.491 & 0.772 \\\\\nDAM-Co-teaching & 0.57 & 0.617 & 0.438 & 0.27 & 0.455 & 0.781 & 0.696 & 0.775 & 0.652 & 0.341 & 0.499 & 0.784 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Evaluation results of co-teaching initialized with different networks.\n
\n \n Douban (Margin) || MAP\n \n Douban (Margin) || MRR\n \n Douban (Margin) || P@1\n \n Douban (Margin) || R10@1\n \n Douban (Margin) || R10@2\n \n Douban (Margin) || R10@5\n \n ECD (Curriculum) || MAP\n \n ECD (Curriculum) || MRR\n \n ECD (Curriculum) || P@1\n \n ECD (Curriculum) || R10@1\n \n ECD (Curriculum) || R10@2\n \n ECD (Curriculum) || R10@5\n
\n SMN-Pre-training\n \n 0.527\n \n 0.57\n \n 0.396\n \n 0.236\n \n 0.392\n \n 0.734\n \n 0.662\n \n 0.742\n \n 0.598\n \n 0.302\n \n 0.464\n \n 0.757\n
\n SMN-Co-teaching\n \n 0.558\n \n 0.602\n \n 0.42\n \n 0.255\n \n 0.431\n \n 0.787\n \n 0.674\n \n 0.765\n \n 0.626\n \n 0.322\n \n 0.485\n \n 0.779\n
\n DAM-Pre-training\n \n 0.552\n \n 0.605\n \n 0.426\n \n 0.258\n \n 0.408\n \n 0.766\n \n 0.685\n \n 0.756\n \n 0.621\n \n 0.325\n \n 0.491\n \n 0.772\n
\n DAM-Co-teaching\n \n 0.57\n \n 0.617\n \n 0.438\n \n 0.27\n \n 0.455\n \n 0.781\n \n 0.696\n \n 0.775\n \n 0.652\n \n 0.341\n \n 0.499\n \n 0.784\n
\n" }, { "instance_id": "9c392e36-2b4d-400f-bc53-9519e5dbc535", "table_id_paper": "P19-1402table_1", "caption": "Performance on the Chimera benchmark dataset with different numbers of context sentences, which is measured by Spearman correlation. Baseline results are from the corresponding papers.", "row_header_level": 2, "row_headers": [ [ "Methods", "Word2vec " ], [ "Methods", "FastText " ], [ "Methods", "Additive " ], [ "Methods", "Additive, no stop words" ], [ "Methods", "nonce2vec " ], [ "Methods", "\u00c3\u00a0 la carte " ], [ "Methods", "HiCE w/o Morph " ], [ "Methods", "HiCE + Morph " ], [ "Methods", "HiCE + Morph + Fine-tune" ], [ "Methods", "HiCE + Morph + MAML" ], [ "Methods", "Oracle Embedding " ] ], "column_header_level": 1, "column_headers": [ [ "2-shot" ], [ "4-shot" ], [ "6-shot" ] ], "contents": [ [ "0.1459", "0.2457", "0.2498" ], [ "0.1775", "0.1738", "0.1294" ], [ "0.3627", "0.3701", "0.3595" ], [ "0.3376", "0.3624", "0.408" ], [ "0.332", "0.3668", "0.389" ], [ "0.3634", "0.3844", "0.3941" ], [ "0.371", "0.3872", "0.4277" ], [ "0.3796", "0.3916", "0.4253" ], [ "0.1403", "0.1837", "0.3145" ], [ "0.3781", "0.4053", "0.4307" ], [ "0.416", "0.4381", "0.4427" ] ], "metrics_loc": "column", "metrics_type": "['correlation', 'correlation', 'correlation']", "target_entity": "['HiCE + Morph + MAML']", "table_html": "
2-shot 4-shot 6-shot
Methods || Word2vec 0.1459 0.2457 0.2498
Methods || FastText 0.1775 0.1738 0.1294
Methods || Additive 0.3627 0.3701 0.3595
Methods || Additive, no stop words 0.3376 0.3624 0.408
Methods || nonce2vec 0.332 0.3668 0.389
Methods || \u00c3\u00a0 la carte 0.3634 0.3844 0.3941
Methods || HiCE w/o Morph 0.371 0.3872 0.4277
Methods || HiCE + Morph 0.3796 0.3916 0.4253
Methods || HiCE + Morph + Fine-tune 0.1403 0.1837 0.3145
Methods || HiCE + Morph + MAML 0.3781 0.4053 0.4307
Methods || Oracle Embedding 0.416 0.4381 0.4427
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P19-1402", "page_no": 6, "dir": "acl2019", "description": "Experimental Results . Table 1 lists the performance of HiCE and baselines with different numbers of context sentences. In particular, our method (HiCE+Morph+MAML)6 achieves the best performance among all the other baseline methods under most settings. Compared with the current state-of-the-art method, `a la carte, the relative improvements (i.e., the performance difference divided by the baseline performance) of HiCE are 4.0%, 5.4% and 9.3% in terms of 2,4,6shot learning, respectively. We also compare our results with that of the oracle embedding, which is the embeddings trained from DT , and used as ground-truth to train HiCE. This results can be regarded as an upper bound. As is shown, when the number of context sentences (K) is relatively large (i.e., K = 6), the performance of HiCE is on a par with the upper bound (Oracle Embedding) and the relative performance difference is merely 2.7%. This indicates the significance of using an advanced aggregation model.", "class_sentence": "[2, 2, 1, 1, 2, 2, 1, 2]", "sentences": "['Experimental Results .', 'Table 1 lists the performance of HiCE and baselines with different numbers of context sentences.', 'In particular, our method (HiCE+Morph+MAML)6 achieves the best performance among all the other baseline methods under most settings.', 'Compared with the current state-of-the-art method, `a la carte, the relative improvements (i.e., the performance difference divided by the baseline performance) of HiCE are 4.0%, 5.4% and 9.3% in terms of 2,4,6shot learning, respectively.', 'We also compare our results with that of the oracle embedding, which is the embeddings trained from DT , and used as ground-truth to train HiCE.', 'This results can be regarded as an upper bound.', 'As is shown, when the number of context sentences (K) is relatively large (i.e., K = 6), the performance of HiCE is on a par with the upper bound (Oracle Embedding) and the relative performance difference is merely 2.7%.', 'This indicates the significance of using an advanced aggregation model.']", "header_mention": "[None, None, ['HiCE + Morph + MAML'], ['HiCE + Morph + MAML'], ['Oracle Embedding '], None, ['HiCE w/o Morph ', 'HiCE + Morph ', 'HiCE + Morph + Fine-tune', 'HiCE + Morph + MAML', 'Oracle Embedding '], None]", "valid": 1, "image_id": "P19-1402-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance on the Chimera benchmark dataset with different numbers of context sentences, which is measured by Spearman correlation. Baseline results are from the corresponding papers.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n 2-shot4-shot6-shot
Methods || Word2vec0.14590.24570.2498
Methods || FastText0.17750.17380.1294
Methods || Additive0.36270.37010.3595
Methods || Additive, no stop words0.33760.36240.408
Methods || nonce2vec0.3320.36680.389
Methods || \u00c3\u00a0 la carte0.36340.38440.3941
Methods || HiCE w/o Morph0.3710.38720.4277
Methods || HiCE + Morph0.37960.39160.4253
Methods || HiCE + Morph + Fine-tune0.14030.18370.3145
Methods || HiCE + Morph + MAML0.37810.40530.4307
Methods || Oracle Embedding0.4160.43810.4427
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance on the Chimera benchmark dataset with different numbers of context sentences, which is measured by Spearman correlation. Baseline results are from the corresponding papers.}\n\\begin{tabular}{lrrr}\n\\toprule\n & 2-shot & 4-shot & 6-shot \\\\\n\\midrule\nMethods $\\parallel$ Word2vec & 0.1459 & 0.2457 & 0.2498 \\\\\nMethods $\\parallel$ FastText & 0.1775 & 0.1738 & 0.1294 \\\\\nMethods $\\parallel$ Additive & 0.3627 & 0.3701 & 0.3595 \\\\\nMethods $\\parallel$ Additive, no stop words & 0.3376 & 0.3624 & 0.408 \\\\\nMethods $\\parallel$ nonce2vec & 0.332 & 0.3668 & 0.389 \\\\\nMethods $\\parallel$ \u00c3 la carte & 0.3634 & 0.3844 & 0.3941 \\\\\nMethods $\\parallel$ HiCE w/o Morph & 0.371 & 0.3872 & 0.4277 \\\\\nMethods $\\parallel$ HiCE + Morph & 0.3796 & 0.3916 & 0.4253 \\\\\nMethods $\\parallel$ HiCE + Morph + Fine-tune & 0.1403 & 0.1837 & 0.3145 \\\\\nMethods $\\parallel$ HiCE + Morph + MAML & 0.3781 & 0.4053 & 0.4307 \\\\\nMethods $\\parallel$ Oracle Embedding & 0.416 & 0.4381 & 0.4427 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Performance on the Chimera benchmark dataset with different numbers of context sentences, which is measured by Spearman correlation. Baseline results are from the corresponding papers.\n
\n \n 2-shot\n \n 4-shot\n \n 6-shot\n
\n Methods || Word2vec\n \n 0.1459\n \n 0.2457\n \n 0.2498\n
\n Methods || FastText\n \n 0.1775\n \n 0.1738\n \n 0.1294\n
\n Methods || Additive\n \n 0.3627\n \n 0.3701\n \n 0.3595\n
\n Methods || Additive, no stop words\n \n 0.3376\n \n 0.3624\n \n 0.408\n
\n Methods || nonce2vec\n \n 0.332\n \n 0.3668\n \n 0.389\n
\n Methods || \u00c3\u00a0 la carte\n \n 0.3634\n \n 0.3844\n \n 0.3941\n
\n Methods || HiCE w/o Morph\n \n 0.371\n \n 0.3872\n \n 0.4277\n
\n Methods || HiCE + Morph\n \n 0.3796\n \n 0.3916\n \n 0.4253\n
\n Methods || HiCE + Morph + Fine-tune\n \n 0.1403\n \n 0.1837\n \n 0.3145\n
\n Methods || HiCE + Morph + MAML\n \n 0.3781\n \n 0.4053\n \n 0.4307\n
\n Methods || Oracle Embedding\n \n 0.416\n \n 0.4381\n \n 0.4427\n
\n" }, { "instance_id": "3dd208c9-16c9-433a-b261-1a5e510b4c3a", "table_id_paper": "P19-1410table_1", "caption": "Discourse segmentation results. Superscript (cid:63) indicates the model is significantly superior to the WLYELMo model with a p-value < 0.01.", "row_header_level": 3, "row_headers": [ [ "Approach", "Human Agreement", "-" ], [ "Approach", "Baselines", "SPADE (Soricut and Marcu, 2003)" ], [ "Approach", "Baselines", "F&R (Fisher and Roark, 2007)" ], [ "Approach", "Baselines", "JCN (Joty et al., 2012)" ], [ "Approach", "Baselines", "SegBotglove (Li et al., 2018)" ], [ "Approach", "Baselines", "WLYELMo (Wang et al., 2018)" ], [ "Approach", "Our Segmenter", "Pointer Net (Glove)" ], [ "Approach", "Our Segmenter", "Pointer Net (BERT)" ], [ "Approach", "Our Segmenter", "Pointer Net (ELMo)" ], [ "Approach", "Our Segmenter", "Pointer Net (ELMo) + Joint training" ] ], "column_header_level": 1, "column_headers": [ [ "Precision" ], [ "Recall" ], [ "F1" ] ], "contents": [ [ "98.5", "98.2", "98.3" ], [ "83.8", "86.8", "85.2" ], [ "91.3", "89.7", "90.5" ], [ "88", "92.3", "90.1" ], [ "91.08\u00b10.46", "91.03 \u00b10.42", "91.05 \u00b10.11" ], [ "92.04 \u00b10.43", "94.41 \u00b10.53", "93.21 \u00b10.33" ], [ "90.55 \u00b10.33", "92.29 \u00b10.09", "91.41 \u00b10.21" ], [ "92.05 \u00b10.44", "95.03 \u00b10.28", "93.51 \u00b10.16" ], [ "94.12\u00b10.20", " 96.63\u00b10.12", " 95.35\u00b10.10" ], [ "93.34\u00b10.23", "97.88\u00b10.16", "95.55\u00b10.13" ] ], "metrics_loc": "column", "metrics_type": "['Precision', 'Recall', 'F1']", "target_entity": "['Our Segmenter']", "table_html": "
Precision Recall F1
Approach || Human Agreement || - 98.5 98.2 98.3
Approach || Baselines || SPADE (Soricut and Marcu, 2003) 83.8 86.8 85.2
Approach || Baselines || F&R (Fisher and Roark, 2007) 91.3 89.7 90.5
Approach || Baselines || JCN (Joty et al., 2012) 88 92.3 90.1
Approach || Baselines || SegBotglove (Li et al., 2018) 91.08\u00b10.46 91.03 \u00b10.42 91.05 \u00b10.11
Approach || Baselines || WLYELMo (Wang et al., 2018) 92.04 \u00b10.43 94.41 \u00b10.53 93.21 \u00b10.33
Approach || Our Segmenter || Pointer Net (Glove) 90.55 \u00b10.33 92.29 \u00b10.09 91.41 \u00b10.21
Approach || Our Segmenter || Pointer Net (BERT) 92.05 \u00b10.44 95.03 \u00b10.28 93.51 \u00b10.16
Approach || Our Segmenter || Pointer Net (ELMo) 94.12\u00b10.20 96.63\u00b10.12 95.35\u00b10.10
Approach || Our Segmenter || Pointer Net (ELMo) + Joint training 93.34\u00b10.23 97.88\u00b10.16 95.55\u00b10.13
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P19-1410", "page_no": 7, "dir": "acl2019", "description": "Results. Table 1 shows our segmentation results. As mentioned in Section 4, we implemented three key improvements on the top of (Li et al., 2018). Using encoder hidden states as decoder inputs and adopting dot product as the attention score function together gives 0.40%-7.29% relative improvement in F1 over the first four baselines. Using ELMo, our segmenter outperforms all the baselines in all three measures. We achieve 2.3%-11.9%, 2.4%-11.3% and 2.3%-12.3% relative improvements inF1, Recall and Precision, respectively. Jointly training with the parser improves this further (95.55 F1). It is worthwhile to mention that our segmenter\u2019s performance of 95.55 F1 is very close to the human agreement of 98.3 F1. ELMo, as a transfer learning method, provides notable improvements. A similar observation was reported in (Wang et al., 2018). Surprisingly, the results with BERT were not as good. We suspect this is due to BERT\u2019s special tokenization.", "class_sentence": "[0, 1, 2, 2, 1, 1, 1, 2, 2, 1, 2]", "sentences": "['Results.', 'Table 1 shows our segmentation results.', 'As mentioned in Section 4, we implemented three key improvements on the top of (Li et al., 2018).', 'Using encoder hidden states as decoder inputs and adopting dot product as the attention score function together gives 0.40%-7.29% relative improvement in F1 over the first four baselines.', 'Using ELMo, our segmenter outperforms all the baselines in all three measures.', 'We achieve 2.3%-11.9%, 2.4%-11.3% and 2.3%-12.3% relative improvements inF1, Recall and Precision, respectively. Jointly training with the parser improves this further (95.55 F1).', 'It is worthwhile to mention that our segmenter\u2019s performance of 95.55 F1 is very close to the human agreement of 98.3 F1.', 'ELMo, as a transfer learning method, provides notable improvements.', 'A similar observation was reported in (Wang et al., 2018).', 'Surprisingly, the results with BERT were not as good.', 'We suspect this is due to BERT\u2019s special tokenization.']", "header_mention": "[None, None, None, None, ['Our Segmenter'], ['Our Segmenter', 'Precision', 'Recall', 'F1'], ['Our Segmenter', 'Human Agreement', 'F1'], ['Pointer Net (ELMo)'], None, ['Pointer Net (BERT)'], ['Pointer Net (BERT)']]", "valid": 1, "image_id": "P19-1410-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Discourse segmentation results. Superscript (cid:63) indicates the model is significantly superior to the WLYELMo model with a p-value < 0.01.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n PrecisionRecallF1
Approach || Human Agreement || -98.598.298.3
Approach || Baselines || SPADE (Soricut and Marcu, 2003)83.886.885.2
Approach || Baselines || F&R (Fisher and Roark, 2007)91.389.790.5
Approach || Baselines || JCN (Joty et al., 2012)8892.390.1
Approach || Baselines || SegBotglove (Li et al., 2018)91.08\u00b10.4691.03 \u00b10.4291.05 \u00b10.11
Approach || Baselines || WLYELMo (Wang et al., 2018)92.04 \u00b10.4394.41 \u00b10.5393.21 \u00b10.33
Approach || Our Segmenter || Pointer Net (Glove)90.55 \u00b10.3392.29 \u00b10.0991.41 \u00b10.21
Approach || Our Segmenter || Pointer Net (BERT)92.05 \u00b10.4495.03 \u00b10.2893.51 \u00b10.16
Approach || Our Segmenter || Pointer Net (ELMo)94.12\u00b10.2096.63\u00b10.1295.35\u00b10.10
Approach || Our Segmenter || Pointer Net (ELMo) + Joint training93.34\u00b10.2397.88\u00b10.1695.55\u00b10.13
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Discourse segmentation results. Superscript (cid:63) indicates the model is significantly superior to the WLYELMo model with a p-value $<$ 0.01.}\n\\begin{tabular}{llll}\n\\toprule\n & Precision & Recall & F1 \\\\\n\\midrule\nApproach $\\parallel$ Human Agreement $\\parallel$ - & 98.5 & 98.2 & 98.3 \\\\\nApproach $\\parallel$ Baselines $\\parallel$ SPADE (Soricut and Marcu, 2003) & 83.8 & 86.8 & 85.2 \\\\\nApproach $\\parallel$ Baselines $\\parallel$ F\\&R (Fisher and Roark, 2007) & 91.3 & 89.7 & 90.5 \\\\\nApproach $\\parallel$ Baselines $\\parallel$ JCN (Joty et al., 2012) & 88 & 92.3 & 90.1 \\\\\nApproach $\\parallel$ Baselines $\\parallel$ SegBotglove (Li et al., 2018) & 91.08$\\pm$0.46 & 91.03 $\\pm$0.42 & 91.05 $\\pm$0.11 \\\\\nApproach $\\parallel$ Baselines $\\parallel$ WLYELMo (Wang et al., 2018) & 92.04 $\\pm$0.43 & 94.41 $\\pm$0.53 & 93.21 $\\pm$0.33 \\\\\nApproach $\\parallel$ Our Segmenter $\\parallel$ Pointer Net (Glove) & 90.55 $\\pm$0.33 & 92.29 $\\pm$0.09 & 91.41 $\\pm$0.21 \\\\\nApproach $\\parallel$ Our Segmenter $\\parallel$ Pointer Net (BERT) & 92.05 $\\pm$0.44 & 95.03 $\\pm$0.28 & 93.51 $\\pm$0.16 \\\\\nApproach $\\parallel$ Our Segmenter $\\parallel$ Pointer Net (ELMo) & 94.12$\\pm$0.20 & 96.63$\\pm$0.12 & 95.35$\\pm$0.10 \\\\\nApproach $\\parallel$ Our Segmenter $\\parallel$ Pointer Net (ELMo) + Joint training & 93.34$\\pm$0.23 & 97.88$\\pm$0.16 & 95.55$\\pm$0.13 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Discourse segmentation results. Superscript (cid:63) indicates the model is significantly superior to the WLYELMo model with a p-value < 0.01.\n
\n \n Precision\n \n Recall\n \n F1\n
\n Approach || Human Agreement || -\n \n 98.5\n \n 98.2\n \n 98.3\n
\n Approach || Baselines || SPADE (Soricut and Marcu, 2003)\n \n 83.8\n \n 86.8\n \n 85.2\n
\n Approach || Baselines || F&R (Fisher and Roark, 2007)\n \n 91.3\n \n 89.7\n \n 90.5\n
\n Approach || Baselines || JCN (Joty et al., 2012)\n \n 88\n \n 92.3\n \n 90.1\n
\n Approach || Baselines || SegBotglove (Li et al., 2018)\n \n 91.08\u00b10.46\n \n 91.03 \u00b10.42\n \n 91.05 \u00b10.11\n
\n Approach || Baselines || WLYELMo (Wang et al., 2018)\n \n 92.04 \u00b10.43\n \n 94.41 \u00b10.53\n \n 93.21 \u00b10.33\n
\n Approach || Our Segmenter || Pointer Net (Glove)\n \n 90.55 \u00b10.33\n \n 92.29 \u00b10.09\n \n 91.41 \u00b10.21\n
\n Approach || Our Segmenter || Pointer Net (BERT)\n \n 92.05 \u00b10.44\n \n 95.03 \u00b10.28\n \n 93.51 \u00b10.16\n
\n Approach || Our Segmenter || Pointer Net (ELMo)\n \n 94.12\u00b10.20\n \n 96.63\u00b10.12\n \n 95.35\u00b10.10\n
\n Approach || Our Segmenter || Pointer Net (ELMo) + Joint training\n \n 93.34\u00b10.23\n \n 97.88\u00b10.16\n \n 95.55\u00b10.13\n
\n" }, { "instance_id": "c0241a6f-d75e-4f01-9f64-419f923a3bbf", "table_id_paper": "P19-1410table_2", "caption": "Parsing results with gold segmentation. Superscript (cid:63) indicates the model is significantly superior to the 2-Stage Parser with a p-value < 0.01.", "row_header_level": 3, "row_headers": [ [ "Approach", "Human Agreement", "-" ], [ "Approach", "Baselines", "SPADE (Soricut and Marcu, 2003)" ], [ "Approach", "Baselines", "DCRF (Joty et al., 2012)" ], [ "Approach", "Baselines", "DPLP (Ji and Eisenstein, 2014)" ], [ "Approach", "Baselines", "2-Stage Parser (Wang et al., 2017)" ], [ "Approach", "Our Parser", "Stack Pointer (ELMo-medium)" ], [ "Approach", "Our Parser", "Stack Pointer (ELMo-large)" ], [ "Approach", "Our Parser", "+ Partial tree information" ], [ "Approach", "Our Parser", "+ Joint training" ] ], "column_header_level": 1, "column_headers": [ [ "Span" ], [ "Nuclearity" ], [ "Relation" ] ], "contents": [ [ "95.7", "90.4", "83" ], [ "93.5", "85.8", "67.6" ], [ "94.6", "86.9", "77.1" ], [ "93.5", "81.3", "70.5" ], [ "95.6", "87.8", "77.6" ], [ "96.37", "89.04*", "79.03*" ], [ "96.86*", "90.77*", "81.12*" ], [ "96.94*", "90.89*", "81.28*" ], [ "97.44*", "91.34*", "81.70*" ] ], "metrics_loc": "column", "metrics_type": "['F1', 'F1', 'F1']", "target_entity": "['Our Parser']", "table_html": "
Span Nuclearity Relation
Approach || Human Agreement || - 95.7 90.4 83
Approach || Baselines || SPADE (Soricut and Marcu, 2003) 93.5 85.8 67.6
Approach || Baselines || DCRF (Joty et al., 2012) 94.6 86.9 77.1
Approach || Baselines || DPLP (Ji and Eisenstein, 2014) 93.5 81.3 70.5
Approach || Baselines || 2-Stage Parser (Wang et al., 2017) 95.6 87.8 77.6
Approach || Our Parser || Stack Pointer (ELMo-medium) 96.37 89.04* 79.03*
Approach || Our Parser || Stack Pointer (ELMo-large) 96.86* 90.77* 81.12*
Approach || Our Parser || + Partial tree information 96.94* 90.89* 81.28*
Approach || Our Parser || + Joint training 97.44* 91.34* 81.70*
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P19-1410", "page_no": 8, "dir": "acl2019", "description": "Thanks to the pointer network as the backbone of our model, we are able to train our segmenter and parser jointly by sharing the same encoder. The last row of Table 2 shows the results when we train the model jointly, and feed the parser with gold EDU segmentation during inference. The performance is improved further with joint training, achieving 97.44, 91.34, 81.70 F1 score, in Span, Nuclearity and Relation, respectively. The results accord with our assumption that discourse segmentation and parsing may benefit from each other. Our parser surpasses human agreement in span and nuclearity. We are also approaching human agreement in the most difficult task of relation labeling. We show a confusion matrix for the relation labels in Figure 5. We see that our model gets confused between relations that are semantically similar (e.g., CAUSE vs. EXPLANATION, COMPARISON vs. CONTRAST, and TEMPORAL vs. JOINT).", "class_sentence": "[2, 1, 1, 2, 1, 2, 0, 0]", "sentences": "['Thanks to the pointer network as the backbone of our model, we are able to train our segmenter and parser jointly by sharing the same encoder.', 'The last row of Table 2 shows the results when we train the model jointly, and feed the parser with gold EDU segmentation during inference.', 'The performance is improved further with joint training, achieving 97.44, 91.34, 81.70 F1 score, in Span, Nuclearity and Relation, respectively.', 'The results accord with our assumption that discourse segmentation and parsing may benefit from each other.', 'Our parser surpasses human agreement in span and nuclearity.', 'We are also approaching human agreement in the most difficult task of relation labeling.', 'We show a confusion matrix for the relation labels in Figure 5.', 'We see that our model gets confused between relations that are semantically similar (e.g., CAUSE vs. EXPLANATION, COMPARISON vs. CONTRAST, and TEMPORAL vs. JOINT).']", "header_mention": "[None, ['+ Joint training'], ['+ Joint training', 'Span', 'Nuclearity', 'Relation'], None, ['Our Parser', 'Human Agreement', 'Span', 'Nuclearity'], None, None, None]", "valid": 1, "image_id": "P19-1410-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Parsing results with gold segmentation. Superscript (cid:63) indicates the model is significantly superior to the 2-Stage Parser with a p-value < 0.01.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n SpanNuclearityRelation
Approach || Human Agreement || -95.790.483
Approach || Baselines || SPADE (Soricut and Marcu, 2003)93.585.867.6
Approach || Baselines || DCRF (Joty et al., 2012)94.686.977.1
Approach || Baselines || DPLP (Ji and Eisenstein, 2014)93.581.370.5
Approach || Baselines || 2-Stage Parser (Wang et al., 2017)95.687.877.6
Approach || Our Parser || Stack Pointer (ELMo-medium)96.3789.04*79.03*
Approach || Our Parser || Stack Pointer (ELMo-large)96.86*90.77*81.12*
Approach || Our Parser || + Partial tree information96.94*90.89*81.28*
Approach || Our Parser || + Joint training97.44*91.34*81.70*
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Parsing results with gold segmentation. Superscript (cid:63) indicates the model is significantly superior to the 2-Stage Parser with a p-value $<$ 0.01.}\n\\begin{tabular}{llll}\n\\toprule\n & Span & Nuclearity & Relation \\\\\n\\midrule\nApproach $\\parallel$ Human Agreement $\\parallel$ - & 95.7 & 90.4 & 83 \\\\\nApproach $\\parallel$ Baselines $\\parallel$ SPADE (Soricut and Marcu, 2003) & 93.5 & 85.8 & 67.6 \\\\\nApproach $\\parallel$ Baselines $\\parallel$ DCRF (Joty et al., 2012) & 94.6 & 86.9 & 77.1 \\\\\nApproach $\\parallel$ Baselines $\\parallel$ DPLP (Ji and Eisenstein, 2014) & 93.5 & 81.3 & 70.5 \\\\\nApproach $\\parallel$ Baselines $\\parallel$ 2-Stage Parser (Wang et al., 2017) & 95.6 & 87.8 & 77.6 \\\\\nApproach $\\parallel$ Our Parser $\\parallel$ Stack Pointer (ELMo-medium) & 96.37 & 89.04* & 79.03* \\\\\nApproach $\\parallel$ Our Parser $\\parallel$ Stack Pointer (ELMo-large) & 96.86* & 90.77* & 81.12* \\\\\nApproach $\\parallel$ Our Parser $\\parallel$ + Partial tree information & 96.94* & 90.89* & 81.28* \\\\\nApproach $\\parallel$ Our Parser $\\parallel$ + Joint training & 97.44* & 91.34* & 81.70* \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Parsing results with gold segmentation. Superscript (cid:63) indicates the model is significantly superior to the 2-Stage Parser with a p-value < 0.01.\n
\n \n Span\n \n Nuclearity\n \n Relation\n
\n Approach || Human Agreement || -\n \n 95.7\n \n 90.4\n \n 83\n
\n Approach || Baselines || SPADE (Soricut and Marcu, 2003)\n \n 93.5\n \n 85.8\n \n 67.6\n
\n Approach || Baselines || DCRF (Joty et al., 2012)\n \n 94.6\n \n 86.9\n \n 77.1\n
\n Approach || Baselines || DPLP (Ji and Eisenstein, 2014)\n \n 93.5\n \n 81.3\n \n 70.5\n
\n Approach || Baselines || 2-Stage Parser (Wang et al., 2017)\n \n 95.6\n \n 87.8\n \n 77.6\n
\n Approach || Our Parser || Stack Pointer (ELMo-medium)\n \n 96.37\n \n 89.04*\n \n 79.03*\n
\n Approach || Our Parser || Stack Pointer (ELMo-large)\n \n 96.86*\n \n 90.77*\n \n 81.12*\n
\n Approach || Our Parser || + Partial tree information\n \n 96.94*\n \n 90.89*\n \n 81.28*\n
\n Approach || Our Parser || + Joint training\n \n 97.44*\n \n 91.34*\n \n 81.70*\n
\n" }, { "instance_id": "29ca5e14-01b1-4aa1-ba02-a09727164be5", "table_id_paper": "P19-1427table_4", "caption": "Results on translating four languages to English for MLE, BLEU, SIMILE and Half. \u2020 denotes statistical significance (p < 0.05) over BLEU and \u2021 denotes statistical significance over MLE. Statistical significance was computed using paired bootstrap resampling.", "row_header_level": 2, "row_headers": [ [ "Model", "MLE" ], [ "Model", "BLEU" ], [ "Model", "SIMILE" ], [ "Model", "Half" ] ], "column_header_level": 2, "column_headers": [ [ "de-en", "BLEU" ], [ "de-en", "SIM" ], [ "cs-en", "BLEU" ], [ "cs-en", "SIM" ], [ "ru-en", "BLEU" ], [ "ru-en", "SIM" ], [ "tr-en", "BLEU" ], [ "tr-en", "SIM" ] ], "contents": [ [ "27.52", "74.96", "17.02", "67.18", "17.92", "70.24", "14.47", "63.52" ], [ "27.95", "86.93", "17.29", "81.92", "17.92", "84.63", "15.00", "80.30" ], [ "28.28", "87.32", "17.51", "82.12", "18.23", "85.12", "15.28", "81.04" ], [ "28.24", "87.11", "17.50", "82.11", "18.24", "85.10", "15.34", "80.64" ] ], "metrics_loc": "column", "metrics_type": "['BLEU', 'SIM', 'BLEU', 'SIM', 'BLEU', 'SIM', 'BLEU', 'SIM']", "target_entity": "['SIMILE']", "table_html": "
de-en || BLEU de-en || SIM cs-en || BLEU cs-en || SIM ru-en || BLEU ru-en || SIM tr-en || BLEU tr-en || SIM
Model || MLE 27.52 74.96 17.02 67.18 17.92 70.24 14.47 63.52
Model || BLEU 27.95 86.93 17.29 81.92 17.92 84.63 15.00 80.30
Model || SIMILE 28.28 87.32 17.51 82.12 18.23 85.12 15.28 81.04
Model || Half 28.24 87.11 17.50 82.11 18.24 85.10 15.34 80.64
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P19-1427", "page_no": 5, "dir": "acl2019", "description": "The results are shown in Table 4. From the table, we see that using SIMILE performs the best when using BLEU and SIM as evaluation metrics for all four languages. It is interesting that using SIMILE in the cost leads to larger BLEU improvements than using BLEU alone, the reasons for which we examine further in the following sections. It is important to emphasize that increasing BLEU was not the goal of our proposed method, human evaluations were our target, but this is a welcome surprise. Similarly, using BLEU as the cost function leads to large gains in SIM, though these gains are not as large as when using SIMILE in training.", "class_sentence": "[1, 1, 1, 2, 1]", "sentences": "['The results are shown in Table 4.', 'From the table, we see that using SIMILE performs the best when using BLEU and SIM as evaluation metrics for all four languages.', 'It is interesting that using SIMILE in the cost leads to larger BLEU improvements than using BLEU alone, the reasons for which we examine further in the following sections.', 'It is important to emphasize that increasing BLEU was not the goal of our proposed method, human evaluations were our target, but this is a welcome surprise.', ' Similarly, using BLEU as the cost function leads to large gains in SIM, though these gains are not as large as when using SIMILE in training.']", "header_mention": "[None, ['SIMILE', 'BLEU', 'SIM'], ['SIMILE', 'BLEU'], ['BLEU'], ['SIM', 'SIMILE']]", "valid": 1, "image_id": "P19-1427-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results on translating four languages to English for MLE, BLEU, SIMILE and Half. \u2020 denotes statistical significance (p < 0.05) over BLEU and \u2021 denotes statistical significance over MLE. Statistical significance was computed using paired bootstrap resampling.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n de-en || BLEUde-en || SIMcs-en || BLEUcs-en || SIMru-en || BLEUru-en || SIMtr-en || BLEUtr-en || SIM
Model || MLE27.5274.9617.0267.1817.9270.2414.4763.52
Model || BLEU27.9586.9317.2981.9217.9284.6315.0080.30
Model || SIMILE28.2887.3217.5182.1218.2385.1215.2881.04
Model || Half28.2487.1117.5082.1118.2485.1015.3480.64
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results on translating four languages to English for MLE, BLEU, SIMILE and Half. $\\dagger$ denotes statistical significance (p $<$ 0.05) over BLEU and $\\ddagger$ denotes statistical significance over MLE. Statistical significance was computed using paired bootstrap resampling.}\n\\begin{tabular}{lrrrrrrrr}\n\\toprule\n & de-en $\\parallel$ BLEU & de-en $\\parallel$ SIM & cs-en $\\parallel$ BLEU & cs-en $\\parallel$ SIM & ru-en $\\parallel$ BLEU & ru-en $\\parallel$ SIM & tr-en $\\parallel$ BLEU & tr-en $\\parallel$ SIM \\\\\n\\midrule\nModel $\\parallel$ MLE & 27.52 & 74.96 & 17.02 & 67.18 & 17.92 & 70.24 & 14.47 & 63.52 \\\\\nModel $\\parallel$ BLEU & 27.95 & 86.93 & 17.29 & 81.92 & 17.92 & 84.63 & 15 & 80.3 \\\\\nModel $\\parallel$ SIMILE & 28.28 & 87.32 & 17.51 & 82.12 & 18.23 & 85.12 & 15.28 & 81.04 \\\\\nModel $\\parallel$ Half & 28.24 & 87.11 & 17.5 & 82.11 & 18.24 & 85.1 & 15.34 & 80.64 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Results on translating four languages to English for MLE, BLEU, SIMILE and Half. \u2020 denotes statistical significance (p < 0.05) over BLEU and \u2021 denotes statistical significance over MLE. Statistical significance was computed using paired bootstrap resampling.\n
\n \n de-en || BLEU\n \n de-en || SIM\n \n cs-en || BLEU\n \n cs-en || SIM\n \n ru-en || BLEU\n \n ru-en || SIM\n \n tr-en || BLEU\n \n tr-en || SIM\n
\n Model || MLE\n \n 27.52\n \n 74.96\n \n 17.02\n \n 67.18\n \n 17.92\n \n 70.24\n \n 14.47\n \n 63.52\n
\n Model || BLEU\n \n 27.95\n \n 86.93\n \n 17.29\n \n 81.92\n \n 17.92\n \n 84.63\n \n 15.00\n \n 80.30\n
\n Model || SIMILE\n \n 28.28\n \n 87.32\n \n 17.51\n \n 82.12\n \n 18.23\n \n 85.12\n \n 15.28\n \n 81.04\n
\n Model || Half\n \n 28.24\n \n 87.11\n \n 17.50\n \n 82.11\n \n 18.24\n \n 85.10\n \n 15.34\n \n 80.64\n
\n" }, { "instance_id": "ab02df71-78b9-4489-801f-fcdf2c6e7a35", "table_id_paper": "P19-1444table_4", "caption": "Ablation study results. Base model means that we does not use schema linking (SL), memory augmented pointer network (MEM) and the coarse-to-fine framework (CF) on IRNet.", "row_header_level": 2, "row_headers": [ [ "Technique", "Base model" ], [ "Technique", "+SL" ], [ "Technique", "+SL + MEM" ], [ "Technique", "+SL + MEM + CF" ] ], "column_header_level": 1, "column_headers": [ [ "IRNet" ], [ "IRNet(BERT)" ] ], "contents": [ [ " 40.5%", "53.90%" ], [ " 48.5%", "60.30%" ], [ " 51.3%", "60.60%" ], [ " 53.2%", "61.90%" ] ], "metrics_loc": "column", "metrics_type": "['F1', 'F1']", "target_entity": "['IRNet', 'IRNet(BERT)']", "table_html": "
IRNet IRNet(BERT)
Technique || Base model 40.5% 53.90%
Technique || +SL 48.5% 60.30%
Technique || +SL + MEM 51.3% 60.60%
Technique || +SL + MEM + CF 53.2% 61.90%
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P19-1444", "page_no": 7, "dir": "acl2019", "description": "Table 4 presents the ablation study results. It is clear that our base model significantly outperforms SyntaxSQLNet, SyntaxSQLNet(augment) and SyntaxSQLNet(BERT). Performing schema linking (\u2018+SL\u2019) brings about 8.5% and 6.4% absolute improvement on IRNet and IRNet(BERT). Predicting columns in the WHERE clause is known to be challenging (Yavuz et al., 2018). The F1 score on the WHERE clause increases by 12.5% when IRNet performs schema linking. The significant improvement demonstrates the effectiveness of schema linking in addressing the lexical problem. Using the memory augmented pointer network (\u2018+MEM\u2019) further improves the performance of IRNet and IRNet(BERT). We observe that the vanilla pointer network is prone to selecting same columns during synthesis. The number of examples suffering from this problem decreases by 70%, when using the memory augmented pointer network. At last, adopting the coarse-to-fine framework (\u2018+CF\u2019) can further boost performance.", "class_sentence": "[1, 2, 1, 2, 2, 2, 1, 2, 2, 1]", "sentences": "['Table 4 presents the ablation study results.', 'It is clear that our base model significantly outperforms SyntaxSQLNet, SyntaxSQLNet(augment) and SyntaxSQLNet(BERT).', 'Performing schema linking (\u2018+SL\u2019) brings about 8.5% and 6.4% absolute improvement on IRNet and IRNet(BERT).', 'Predicting columns in the WHERE clause is known to be challenging (Yavuz et al., 2018).', 'The F1 score on the WHERE clause increases by 12.5% when IRNet performs schema linking.', 'The significant improvement demonstrates the effectiveness of schema linking in addressing the lexical problem.', 'Using the memory augmented pointer network (\u2018+MEM\u2019) further improves the performance of IRNet and IRNet(BERT).', 'We observe that the vanilla pointer network is prone to selecting same columns during synthesis.', 'The number of examples suffering from this problem decreases by 70%, when using the memory augmented pointer network.', 'At last, adopting the coarse-to-fine framework (\u2018+CF\u2019) can further boost performance.']", "header_mention": "[None, None, ['+SL', 'IRNet', 'IRNet(BERT)'], None, None, None, ['+SL + MEM', 'IRNet', 'IRNet(BERT)'], None, None, ['+SL + MEM + CF']]", "valid": 1, "image_id": "P19-1444-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Ablation study results. Base model means that we does not use schema linking (SL), memory augmented pointer network (MEM) and the coarse-to-fine framework (CF) on IRNet.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n IRNetIRNet(BERT)
Technique || Base model40.5%53.90%
Technique || +SL48.5%60.30%
Technique || +SL + MEM51.3%60.60%
Technique || +SL + MEM + CF53.2%61.90%
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Ablation study results. Base model means that we does not use schema linking (SL), memory augmented pointer network (MEM) and the coarse-to-fine framework (CF) on IRNet.}\n\\begin{tabular}{lll}\n\\toprule\n & IRNet & IRNet(BERT) \\\\\n\\midrule\nTechnique $\\parallel$ Base model & 40.5\\% & 53.90\\% \\\\\nTechnique $\\parallel$ +SL & 48.5\\% & 60.30\\% \\\\\nTechnique $\\parallel$ +SL + MEM & 51.3\\% & 60.60\\% \\\\\nTechnique $\\parallel$ +SL + MEM + CF & 53.2\\% & 61.90\\% \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Ablation study results. Base model means that we does not use schema linking (SL), memory augmented pointer network (MEM) and the coarse-to-fine framework (CF) on IRNet.\n
\n \n IRNet\n \n IRNet(BERT)\n
\n Technique || Base model\n \n 40.5%\n \n 53.90%\n
\n Technique || +SL\n \n 48.5%\n \n 60.30%\n
\n Technique || +SL + MEM\n \n 51.3%\n \n 60.60%\n
\n Technique || +SL + MEM + CF\n \n 53.2%\n \n 61.90%\n
\n" }, { "instance_id": "c234da36-c0ea-4636-90c5-a0410b738289", "table_id_paper": "P19-1458table_5", "caption": "Low-shot learning results for the Yahoo dataset, in error percentages. Transfer learning-based 3 of the total dataset, while the methods are trained on 1 other models are trained on the whole dataset.", "row_header_level": 2, "row_headers": [ [ "Model", "RNN Baseline" ], [ "Model", "LSTM Baseline" ], [ "Model", "KimCNN Kim (2014)" ], [ "Model", "Self Attention Lin et al. (2017)" ], [ "Model", "RCNN Lai et al. (2015)" ], [ "Model", "BCN+ELMo" ], [ "Model", "ULMFiT" ], [ "Model", "BERTBASE" ] ], "column_header_level": 1, "column_headers": [ [ "Yahoo Binary" ] ], "contents": [ [ "35.29" ], [ "32.41" ], [ "14.25" ], [ "13.16" ], [ "12.67" ], [ "13.51" ], [ "10.62" ], [ "10.14" ] ], "metrics_loc": "column", "metrics_type": "['error']", "target_entity": "['ULMFiT', 'BERTBASE', 'BCN+ELMo']", "table_html": "
Yahoo Binary
Model || RNN Baseline 35.29
Model || LSTM Baseline 32.41
Model || KimCNN Kim (2014) 14.25
Model || Self Attention Lin et al. (2017) 13.16
Model || RCNN Lai et al. (2015) 12.67
Model || BCN+ELMo 13.51
Model || ULMFiT 10.62
Model || BERTBASE 10.14
", "table_name": "Table 5", "table_id": "table_5", "paper_id": "P19-1458", "page_no": 4, "dir": "acl2019", "description": "7.2 Low-Shot Learning . Low-shot learning refers to the practice of feeding a model with a small amount of training data, contrary to the normal practice of using a large amount of data. We chose the Yahoo dataset for this experiment due to its small size. Experimental results in Table 5 show that, with only 1 3 of the total dataset, ULMFiT and BERT perform better than task-specific models, while BCN+ELMo shows a comparable result. Clearly, this shows that the models have learned significantly during the transfer learning process.", "class_sentence": "[2, 2, 2, 1, 1]", "sentences": "['7.2 Low-Shot Learning .', 'Low-shot learning refers to the practice of feeding a model with a small amount of training data, contrary to the normal practice of using a large amount of data.', 'We chose the Yahoo dataset for this experiment due to its small size.', 'Experimental results in Table 5 show that, with only 1 3 of the total dataset, ULMFiT and BERT perform better than task-specific models, while BCN+ELMo shows a comparable result.', 'Clearly, this shows that the models have learned significantly during the transfer learning process.']", "header_mention": "[None, None, ['Yahoo Binary'], ['ULMFiT', 'BERTBASE', 'BCN+ELMo'], None]", "valid": 1, "image_id": "P19-1458-Table5-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Low-shot learning results for the Yahoo dataset, in error percentages. Transfer learning-based 3 of the total dataset, while the methods are trained on 1 other models are trained on the whole dataset.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Yahoo Binary
Model || RNN Baseline35.29
Model || LSTM Baseline32.41
Model || KimCNN Kim (2014)14.25
Model || Self Attention Lin et al. (2017)13.16
Model || RCNN Lai et al. (2015)12.67
Model || BCN+ELMo13.51
Model || ULMFiT10.62
Model || BERTBASE10.14
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{4}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Low-shot learning results for the Yahoo dataset, in error percentages. Transfer learning-based 3 of the total dataset, while the methods are trained on 1 other models are trained on the whole dataset.}\n\\begin{tabular}{lr}\n\\toprule\n & Yahoo Binary \\\\\n\\midrule\nModel $\\parallel$ RNN Baseline & 35.29 \\\\\nModel $\\parallel$ LSTM Baseline & 32.41 \\\\\nModel $\\parallel$ KimCNN Kim (2014) & 14.25 \\\\\nModel $\\parallel$ Self Attention Lin et al. (2017) & 13.16 \\\\\nModel $\\parallel$ RCNN Lai et al. (2015) & 12.67 \\\\\nModel $\\parallel$ BCN+ELMo & 13.51 \\\\\nModel $\\parallel$ ULMFiT & 10.62 \\\\\nModel $\\parallel$ BERTBASE & 10.14 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 5: Low-shot learning results for the Yahoo dataset, in error percentages. Transfer learning-based 3 of the total dataset, while the methods are trained on 1 other models are trained on the whole dataset.\n
\n \n Yahoo Binary\n
\n Model || RNN Baseline\n \n 35.29\n
\n Model || LSTM Baseline\n \n 32.41\n
\n Model || KimCNN Kim (2014)\n \n 14.25\n
\n Model || Self Attention Lin et al. (2017)\n \n 13.16\n
\n Model || RCNN Lai et al. (2015)\n \n 12.67\n
\n Model || BCN+ELMo\n \n 13.51\n
\n Model || ULMFiT\n \n 10.62\n
\n Model || BERTBASE\n \n 10.14\n
\n" }, { "instance_id": "c207db98-79ab-485c-b2e0-b17ef44ade70", "table_id_paper": "P19-1475table_2", "caption": "Results on recast MNLI and JOCI.", "row_header_level": 2, "row_headers": [ [ "Dataset", "MNLI 1" ], [ "Dataset", "MNLI 2" ], [ "Dataset", "JOCI 1" ], [ "Dataset", "JOCI 2" ] ], "column_header_level": 1, "column_headers": [ [ "Log loss" ], [ "Margin loss" ] ], "contents": [ [ "93.6", "93.4" ], [ "87.9", "87.9" ], [ "86.6", "86.9" ], [ "76.6", "78" ] ], "metrics_loc": "column", "metrics_type": "['Log loss', 'Margin loss']", "target_entity": "['MNLI 1', 'MNLI 2', 'JOCI 1', 'JOCI 2']", "table_html": "
Log loss Margin loss
Dataset || MNLI 1 93.6 93.4
Dataset || MNLI 2 87.9 87.9
Dataset || JOCI 1 86.6 86.9
Dataset || JOCI 2 76.6 78
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P19-1475", "page_no": 3, "dir": "acl2019", "description": "Results on Recast MNLI and JOCI. Table 2 shows results on the recast MNLI and JOCI datasets. We find that for the two synthetic MNLI datasets, margin-loss performs similarly to cross entropy log-loss. Shifting to the JOCI datasets, with less extreme (contradiction / entailed) hypotheses, especially in the adversarial JOCI2 variant, marginloss outperforms log-loss.", "class_sentence": "[2, 1, 1, 1]", "sentences": "['Results on Recast MNLI and JOCI.', 'Table 2 shows results on the recast MNLI and JOCI datasets.', 'We find that for the two synthetic MNLI datasets, margin-loss performs similarly to cross entropy log-loss.', 'Shifting to the JOCI datasets, with less extreme (contradiction / entailed) hypotheses, especially in the adversarial JOCI2 variant, marginloss outperforms log-loss.']", "header_mention": "[None, ['MNLI 1', 'JOCI 1', 'MNLI 2', 'JOCI 2'], ['MNLI 1', 'MNLI 2', 'Margin loss', 'Log loss'], ['JOCI 1', 'JOCI 2', 'Margin loss', 'Log loss']]", "valid": 1, "image_id": "P19-1475-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results on recast MNLI and JOCI.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Log lossMargin loss
Dataset || MNLI 193.693.4
Dataset || MNLI 287.987.9
Dataset || JOCI 186.686.9
Dataset || JOCI 276.678
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results on recast MNLI and JOCI.}\n\\begin{tabular}{lrr}\n\\toprule\n & Log loss & Margin loss \\\\\n\\midrule\nDataset $\\parallel$ MNLI 1 & 93.6 & 93.4 \\\\\nDataset $\\parallel$ MNLI 2 & 87.9 & 87.9 \\\\\nDataset $\\parallel$ JOCI 1 & 86.6 & 86.9 \\\\\nDataset $\\parallel$ JOCI 2 & 76.6 & 78 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Results on recast MNLI and JOCI.\n
\n \n Log loss\n \n Margin loss\n
\n Dataset || MNLI 1\n \n 93.6\n \n 93.4\n
\n Dataset || MNLI 2\n \n 87.9\n \n 87.9\n
\n Dataset || JOCI 1\n \n 86.6\n \n 86.9\n
\n Dataset || JOCI 2\n \n 76.6\n \n 78\n
\n" }, { "instance_id": "1d30a781-ba63-4f83-8836-6fc46313d6b1", "table_id_paper": "P19-1475table_3", "caption": "Experimental results on COPA test set.", "row_header_level": 2, "row_headers": [ [ "Method", "PMI (Jabeen et al., 2014)" ], [ "Method", "PMI EX (Gordon et al., 2011)" ], [ "Method", "CS (Luo et al., 2016)" ], [ "Method", "CS MWP (Sasaki et al., 2017)" ], [ "Method", "BERTlog (ours)" ], [ "Method", "BERTmargin (ours)" ] ], "column_header_level": 1, "column_headers": [ [ "Acc (%)" ] ], "contents": [ [ "58.8" ], [ "65.4" ], [ "70.2" ], [ "71.2" ], [ "73.4" ], [ "75.4" ] ], "metrics_loc": "column", "metrics_type": "['Acc (%)']", "target_entity": "['BERTmargin (ours)', 'BERTlog (ours)']", "table_html": "
Acc (%)
Method || PMI (Jabeen et al., 2014) 58.8
Method || PMI EX (Gordon et al., 2011) 65.4
Method || CS (Luo et al., 2016) 70.2
Method || CS MWP (Sasaki et al., 2017) 71.2
Method || BERTlog (ours) 73.4
Method || BERTmargin (ours) 75.4
", "table_name": "Table 3", "table_id": "table_3", "paper_id": "P19-1475", "page_no": 3, "dir": "acl2019", "description": "Results on COPA. Table 3 shows our results on COPA. Compared with previous state-of-theart knowledge-driven baseline methods, a BERT model trained with a log-loss achieves better performance. When training the BERT model with a margin-loss instead of a log-loss, our method gets the new state-of-the-art result on the established COPA splits, with an accuracy of 75.4%.3 Analyses Table 4 shows some examples from the MNLI1, JOCI1 and COPA datasets, with scores 3 We exclude a blog-posted GPT result, which comes.", "class_sentence": "[2, 1, 1, 1]", "sentences": "['Results on COPA.', 'Table 3 shows our results on COPA.', 'Compared with previous state-of-theart knowledge-driven baseline methods, a BERT model trained with a log-loss achieves better performance.', 'When training the BERT model with a margin-loss instead of a log-loss, our method gets the new state-of-the-art result on the established COPA splits, with an accuracy of 75.4%.']", "header_mention": "[None, None, ['PMI (Jabeen et al., 2014)', 'PMI EX (Gordon et al., 2011)', 'CS (Luo et al., 2016)', 'CS MWP (Sasaki et al., 2017)', 'BERTlog (ours)'], ['BERTmargin (ours)', 'BERTlog (ours)', 'Acc (%)']]", "valid": 1, "image_id": "P19-1475-Table3-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Experimental results on COPA test set.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Acc (%)
Method || PMI (Jabeen et al., 2014)58.8
Method || PMI EX (Gordon et al., 2011)65.4
Method || CS (Luo et al., 2016)70.2
Method || CS MWP (Sasaki et al., 2017)71.2
Method || BERTlog (ours)73.4
Method || BERTmargin (ours)75.4
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{2}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Experimental results on COPA test set.}\n\\begin{tabular}{lr}\n\\toprule\n & Acc (\\%) \\\\\n\\midrule\nMethod $\\parallel$ PMI (Jabeen et al., 2014) & 58.8 \\\\\nMethod $\\parallel$ PMI EX (Gordon et al., 2011) & 65.4 \\\\\nMethod $\\parallel$ CS (Luo et al., 2016) & 70.2 \\\\\nMethod $\\parallel$ CS MWP (Sasaki et al., 2017) & 71.2 \\\\\nMethod $\\parallel$ BERTlog (ours) & 73.4 \\\\\nMethod $\\parallel$ BERTmargin (ours) & 75.4 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 3: Experimental results on COPA test set.\n
\n \n Acc (%)\n
\n Method || PMI (Jabeen et al., 2014)\n \n 58.8\n
\n Method || PMI EX (Gordon et al., 2011)\n \n 65.4\n
\n Method || CS (Luo et al., 2016)\n \n 70.2\n
\n Method || CS MWP (Sasaki et al., 2017)\n \n 71.2\n
\n Method || BERTlog (ours)\n \n 73.4\n
\n Method || BERTmargin (ours)\n \n 75.4\n
\n" }, { "instance_id": "87d057e0-ac1f-45e1-8cb3-be8c7a2df47f", "table_id_paper": "P19-1487table_2", "caption": "Results on CQA dev-random-split with CoS-E used during training.", "row_header_level": 2, "row_headers": [ [ "Method", "BERT (baseline)" ], [ "Method", "CoS-E-open-ended" ], [ "Method", "CAGE-reasoning" ] ], "column_header_level": 1, "column_headers": [ [ "Accuracy (%)" ] ], "contents": [ [ "63.8" ], [ "65.5" ], [ "72.6" ] ], "metrics_loc": "column", "metrics_type": "['Accuracy (%)']", "target_entity": "['CAGE-reasoning']", "table_html": "
Accuracy (%)
Method || BERT (baseline) 63.8
Method || CoS-E-open-ended 65.5
Method || CAGE-reasoning 72.6
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P19-1487", "page_no": 6, "dir": "acl2019", "description": "Table 2 shows results that compare a BERT baseline that uses only the CQA inputs and the same architecture but trained using inputs that contain explanations from CoS-E during training. The BERT baseline model reaches 64% accuracy and adding open-ended human explanations (CoS-E-open-ended) alongside the questions during training results in a 2% boost in accuracy. By generating explanations as described in Section 4.1, we can give the commonsense question answering model access to an explanation that is not conditioned on the ground truth. These explanations (CAGE-reasoning) can be provided during both training and validation and increases the accuracy to 72%.", "class_sentence": "[1, 1, 2, 1]", "sentences": "['Table 2 shows results that compare a BERT baseline that uses only the CQA inputs and the same architecture but trained using inputs that contain explanations from CoS-E during training.', 'The BERT baseline model reaches 64% accuracy and adding open-ended human explanations (CoS-E-open-ended) alongside the questions during training results in a 2% boost in accuracy.', 'By generating explanations as described in Section 4.1, we can give the commonsense question answering model access to an explanation that is not conditioned on the ground truth.', 'These explanations (CAGE-reasoning) can be provided during both training and validation and increases the accuracy to 72%.']", "header_mention": "[None, ['BERT (baseline)', 'Accuracy (%)', 'CoS-E-open-ended'], None, ['CAGE-reasoning']]", "valid": 1, "image_id": "P19-1487-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results on CQA dev-random-split with CoS-E used during training.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Accuracy (%)
Method || BERT (baseline)63.8
Method || CoS-E-open-ended65.5
Method || CAGE-reasoning72.6
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results on CQA dev-random-split with CoS-E used during training.}\n\\begin{tabular}{lr}\n\\toprule\n & Accuracy (\\%) \\\\\n\\midrule\nMethod $\\parallel$ BERT (baseline) & 63.8 \\\\\nMethod $\\parallel$ CoS-E-open-ended & 65.5 \\\\\nMethod $\\parallel$ CAGE-reasoning & 72.6 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Results on CQA dev-random-split with CoS-E used during training.\n
\n \n Accuracy (%)\n
\n Method || BERT (baseline)\n \n 63.8\n
\n Method || CoS-E-open-ended\n \n 65.5\n
\n Method || CAGE-reasoning\n \n 72.6\n
\n" }, { "instance_id": "4e3aa65e-3faa-48a4-a95a-bd2186876fcb", "table_id_paper": "P19-1530table_2", "caption": "Comparison for nonlocal dependency identification on the test data.", "row_header_level": 1, "row_headers": [ [ "(Johnson, 2002)" ], [ "(Dienes and Dubey, 2003)" ], [ "(Campbell, 2004)" ], [ "(Schmid, 2006)" ], [ "(Cai et al., 2011)" ], [ "(Hayashi and Nagata, 2016)" ], [ "(Kato and Matsubara, 2016)" ], [ "(Kummerfeld and Klein, 2017)" ], [ "(Johnson, 2002)" ], [ "(Campbell, 2004)" ], [ "ours" ], [ "ours (with ELMo)" ], [ "ours (with BERT)" ] ], "column_header_level": 2, "column_headers": [ [ "Empty element detection (Fillers are ignored.)", "pre." ], [ "Empty element detection (Fillers are ignored.)", "rec." ], [ "Empty element detection (Fillers are ignored.)", "F1" ], [ "Nonlocal dependency identification", "pre." ], [ "Nonlocal dependency identification", "rec." ], [ "Nonlocal dependency identification", "F1" ], [ "Nonlocal dependency identification (Unindexed empty elements are excluded.)", "pre." ], [ "Nonlocal dependency identification (Unindexed empty elements are excluded.)", "rec." ], [ "Nonlocal dependency identification (Unindexed empty elements are excluded.)", "F1" ] ], "contents": [ [ "85", "74", "79", "73", "63", "68", "\u2013", "\u2013", "\u2013" ], [ "\u2013", "\u2013", "\u2013", "81.5", "68.7", "74.6", "\u2013", "\u2013", "\u2013" ], [ "85.2", "81.7", "83.4", "78.3", "75.1", "76.7", "\u2013", "\u2013", "\u2013" ], [ "86", "82.3", "84.1", "\u2013", "\u2013", "\u2013", "81.7", "73.5", "77.4" ], [ "90.1", "79.5", "84.5", "\u2013", "\u2013", "\u2013", "\u2013", "\u2013", "\u2013" ], [ "90.3", "81.7", "85.8", "\u2013", "\u2013", "\u2013", "\u2013", "\u2013", "\u2013" ], [ "88.5", "82.1", "85.2", "81.4", "75.5", "78.4", "79.8", "73.8", "76.7" ], [ "89.5", "81.6", "85.4", "74.3", "67.3", "70.6", "\u2013", "\u2013", "\u2013" ], [ "93", "83", "88", "80", "70", "75", "\u2013", "\u2013", "\u2013" ], [ "94.9", "91.1", "93", "90.1", "86.6", "88.4", "\u2013", "\u2013", "\u2013" ], [ "92.6", "87.7", "90.1", "88.1", "83.4", "85.7", "88.4", "81.1", "84.6" ], [ "94.2", "90.3", "92.3", "89.9", "86.2", "88", "90.4", "84.1", "87.2" ], [ "94.9", "91.4", "93.1", "90.8", "87.4", "89", "91.6", "84.9", "88.1" ] ], "metrics_loc": "column", "metrics_type": "['pre.', 'rec.', 'F1', 'pre.', 'rec.', 'F1', 'pre.', 'rec.', 'F1']", "target_entity": "['ours']", "table_html": "
Empty element detection (Fillers are ignored.) || pre. Empty element detection (Fillers are ignored.) || rec. Empty element detection (Fillers are ignored.) || F1 Nonlocal dependency identification || pre. Nonlocal dependency identification || rec. Nonlocal dependency identification || F1 Nonlocal dependency identification (Unindexed empty elements are excluded.) || pre. Nonlocal dependency identification (Unindexed empty elements are excluded.) || rec. Nonlocal dependency identification (Unindexed empty elements are excluded.) || F1
(Johnson, 2002) 85 74 79 73 63 68 \u2013 \u2013 \u2013
(Dienes and Dubey, 2003) \u2013 \u2013 \u2013 81.5 68.7 74.6 \u2013 \u2013 \u2013
(Campbell, 2004) 85.2 81.7 83.4 78.3 75.1 76.7 \u2013 \u2013 \u2013
(Schmid, 2006) 86 82.3 84.1 \u2013 \u2013 \u2013 81.7 73.5 77.4
(Cai et al., 2011) 90.1 79.5 84.5 \u2013 \u2013 \u2013 \u2013 \u2013 \u2013
(Hayashi and Nagata, 2016) 90.3 81.7 85.8 \u2013 \u2013 \u2013 \u2013 \u2013 \u2013
(Kato and Matsubara, 2016) 88.5 82.1 85.2 81.4 75.5 78.4 79.8 73.8 76.7
(Kummerfeld and Klein, 2017) 89.5 81.6 85.4 74.3 67.3 70.6 \u2013 \u2013 \u2013
(Johnson, 2002) 93 83 88 80 70 75 \u2013 \u2013 \u2013
(Campbell, 2004) 94.9 91.1 93 90.1 86.6 88.4 \u2013 \u2013 \u2013
ours 92.6 87.7 90.1 88.1 83.4 85.7 88.4 81.1 84.6
ours (with ELMo) 94.2 90.3 92.3 89.9 86.2 88 90.4 84.1 87.2
ours (with BERT) 94.9 91.4 93.1 90.8 87.4 89 91.6 84.9 88.1
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P19-1530", "page_no": 4, "dir": "acl2019", "description": "Table 2 summarizes the performances of our system and previous ones. These results demonstrate that our system significantly outperforms the previous methods in nonlocal dependency identification. Although the main reason for this is because of the performance of the K&K parser, the important point is that our proposed approximation method enables us to use the K&K parser for the nonlocal dependency identification task. The previous methods that introduce additional operations cannot adopt such parser directly. On the other hand, although post-processing approach can use any parser in pre-processing, our approach outperforms the post-processing approach, even if the pre-processing parser is assumed to always generate gold PTB trees.", "class_sentence": "[1, 1, 2, 2, 1]", "sentences": "['Table 2 summarizes the performances of our system and previous ones.', 'These results demonstrate that our system significantly outperforms the previous methods in nonlocal dependency identification.', 'Although the main reason for this is because of the performance of the K&K parser, the important point is that our proposed approximation method enables us to use the K&K parser for the nonlocal dependency identification task.', 'The previous methods that introduce additional operations cannot adopt such parser directly.', 'On the other hand, although post-processing approach can use any parser in pre-processing, our approach outperforms the post-processing approach, even if the pre-processing parser is assumed to always generate gold PTB trees.']", "header_mention": "[None, ['ours'], ['ours'], None, None]", "valid": 1, "image_id": "P19-1530-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison for nonlocal dependency identification on the test data.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Empty element detection (Fillers are ignored.) || pre.Empty element detection (Fillers are ignored.) || rec.Empty element detection (Fillers are ignored.) || F1Nonlocal dependency identification || pre.Nonlocal dependency identification || rec.Nonlocal dependency identification || F1Nonlocal dependency identification (Unindexed empty elements are excluded.) || pre.Nonlocal dependency identification (Unindexed empty elements are excluded.) || rec.Nonlocal dependency identification (Unindexed empty elements are excluded.) || F1
(Johnson, 2002)857479736368\u2013\u2013\u2013
(Dienes and Dubey, 2003)\u2013\u2013\u201381.568.774.6\u2013\u2013\u2013
(Campbell, 2004)85.281.783.478.375.176.7\u2013\u2013\u2013
(Schmid, 2006)8682.384.1\u2013\u2013\u201381.773.577.4
(Cai et al., 2011)90.179.584.5\u2013\u2013\u2013\u2013\u2013\u2013
(Hayashi and Nagata, 2016)90.381.785.8\u2013\u2013\u2013\u2013\u2013\u2013
(Kato and Matsubara, 2016)88.582.185.281.475.578.479.873.876.7
(Kummerfeld and Klein, 2017)89.581.685.474.367.370.6\u2013\u2013\u2013
(Johnson, 2002)938388807075\u2013\u2013\u2013
(Campbell, 2004)94.991.19390.186.688.4\u2013\u2013\u2013
ours92.687.790.188.183.485.788.481.184.6
ours (with ELMo)94.290.392.389.986.28890.484.187.2
ours (with BERT)94.991.493.190.887.48991.684.988.1
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison for nonlocal dependency identification on the test data.}\n\\begin{tabular}{llllllllll}\n\\toprule\n & Empty element detection (Fillers are ignored.) $\\parallel$ pre. & Empty element detection (Fillers are ignored.) $\\parallel$ rec. & Empty element detection (Fillers are ignored.) $\\parallel$ F1 & Nonlocal dependency identification $\\parallel$ pre. & Nonlocal dependency identification $\\parallel$ rec. & Nonlocal dependency identification $\\parallel$ F1 & Nonlocal dependency identification (Unindexed empty elements are excluded.) $\\parallel$ pre. & Nonlocal dependency identification (Unindexed empty elements are excluded.) $\\parallel$ rec. & Nonlocal dependency identification (Unindexed empty elements are excluded.) $\\parallel$ F1 \\\\\n\\midrule\n(Johnson, 2002) & 85 & 74 & 79 & 73 & 63 & 68 & \u2013 & \u2013 & \u2013 \\\\\n(Dienes and Dubey, 2003) & \u2013 & \u2013 & \u2013 & 81.5 & 68.7 & 74.6 & \u2013 & \u2013 & \u2013 \\\\\n(Campbell, 2004) & 85.2 & 81.7 & 83.4 & 78.3 & 75.1 & 76.7 & \u2013 & \u2013 & \u2013 \\\\\n(Schmid, 2006) & 86 & 82.3 & 84.1 & \u2013 & \u2013 & \u2013 & 81.7 & 73.5 & 77.4 \\\\\n(Cai et al., 2011) & 90.1 & 79.5 & 84.5 & \u2013 & \u2013 & \u2013 & \u2013 & \u2013 & \u2013 \\\\\n(Hayashi and Nagata, 2016) & 90.3 & 81.7 & 85.8 & \u2013 & \u2013 & \u2013 & \u2013 & \u2013 & \u2013 \\\\\n(Kato and Matsubara, 2016) & 88.5 & 82.1 & 85.2 & 81.4 & 75.5 & 78.4 & 79.8 & 73.8 & 76.7 \\\\\n(Kummerfeld and Klein, 2017) & 89.5 & 81.6 & 85.4 & 74.3 & 67.3 & 70.6 & \u2013 & \u2013 & \u2013 \\\\\n(Johnson, 2002) & 93 & 83 & 88 & 80 & 70 & 75 & \u2013 & \u2013 & \u2013 \\\\\n(Campbell, 2004) & 94.9 & 91.1 & 93 & 90.1 & 86.6 & 88.4 & \u2013 & \u2013 & \u2013 \\\\\nours & 92.6 & 87.7 & 90.1 & 88.1 & 83.4 & 85.7 & 88.4 & 81.1 & 84.6 \\\\\nours (with ELMo) & 94.2 & 90.3 & 92.3 & 89.9 & 86.2 & 88 & 90.4 & 84.1 & 87.2 \\\\\nours (with BERT) & 94.9 & 91.4 & 93.1 & 90.8 & 87.4 & 89 & 91.6 & 84.9 & 88.1 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Comparison for nonlocal dependency identification on the test data.\n
\n \n Empty element detection (Fillers are ignored.) || pre.\n \n Empty element detection (Fillers are ignored.) || rec.\n \n Empty element detection (Fillers are ignored.) || F1\n \n Nonlocal dependency identification || pre.\n \n Nonlocal dependency identification || rec.\n \n Nonlocal dependency identification || F1\n \n Nonlocal dependency identification (Unindexed empty elements are excluded.) || pre.\n \n Nonlocal dependency identification (Unindexed empty elements are excluded.) || rec.\n \n Nonlocal dependency identification (Unindexed empty elements are excluded.) || F1\n
\n (Johnson, 2002)\n \n 85\n \n 74\n \n 79\n \n 73\n \n 63\n \n 68\n \n \u2013\n \n \u2013\n \n \u2013\n
\n (Dienes and Dubey, 2003)\n \n \u2013\n \n \u2013\n \n \u2013\n \n 81.5\n \n 68.7\n \n 74.6\n \n \u2013\n \n \u2013\n \n \u2013\n
\n (Campbell, 2004)\n \n 85.2\n \n 81.7\n \n 83.4\n \n 78.3\n \n 75.1\n \n 76.7\n \n \u2013\n \n \u2013\n \n \u2013\n
\n (Schmid, 2006)\n \n 86\n \n 82.3\n \n 84.1\n \n \u2013\n \n \u2013\n \n \u2013\n \n 81.7\n \n 73.5\n \n 77.4\n
\n (Cai et al., 2011)\n \n 90.1\n \n 79.5\n \n 84.5\n \n \u2013\n \n \u2013\n \n \u2013\n \n \u2013\n \n \u2013\n \n \u2013\n
\n (Hayashi and Nagata, 2016)\n \n 90.3\n \n 81.7\n \n 85.8\n \n \u2013\n \n \u2013\n \n \u2013\n \n \u2013\n \n \u2013\n \n \u2013\n
\n (Kato and Matsubara, 2016)\n \n 88.5\n \n 82.1\n \n 85.2\n \n 81.4\n \n 75.5\n \n 78.4\n \n 79.8\n \n 73.8\n \n 76.7\n
\n (Kummerfeld and Klein, 2017)\n \n 89.5\n \n 81.6\n \n 85.4\n \n 74.3\n \n 67.3\n \n 70.6\n \n \u2013\n \n \u2013\n \n \u2013\n
\n (Johnson, 2002)\n \n 93\n \n 83\n \n 88\n \n 80\n \n 70\n \n 75\n \n \u2013\n \n \u2013\n \n \u2013\n
\n (Campbell, 2004)\n \n 94.9\n \n 91.1\n \n 93\n \n 90.1\n \n 86.6\n \n 88.4\n \n \u2013\n \n \u2013\n \n \u2013\n
\n ours\n \n 92.6\n \n 87.7\n \n 90.1\n \n 88.1\n \n 83.4\n \n 85.7\n \n 88.4\n \n 81.1\n \n 84.6\n
\n ours (with ELMo)\n \n 94.2\n \n 90.3\n \n 92.3\n \n 89.9\n \n 86.2\n \n 88\n \n 90.4\n \n 84.1\n \n 87.2\n
\n ours (with BERT)\n \n 94.9\n \n 91.4\n \n 93.1\n \n 90.8\n \n 87.4\n \n 89\n \n 91.6\n \n 84.9\n \n 88.1\n
\n" }, { "instance_id": "d70f495d-76e4-4ebc-b554-9677cf021ded", "table_id_paper": "P19-1565table_4", "caption": "Results of Self-Play Evaluation.", "row_header_level": 2, "row_headers": [ [ "System", "Retrieval" ], [ "System", "Retrieval-Stgy" ], [ "System", "Ours-PMI" ], [ "System", "Ours-Neural" ], [ "System", "Ours-Kernel" ] ], "column_header_level": 1, "column_headers": [ [ "Succ. (%)" ], [ "#Turns" ] ], "contents": [ [ "9.8", "3.26" ], [ "67.2", "6.56" ], [ "47.4", "5.12" ], [ "51.6", "4.29" ], [ "75", "4.2" ] ], "metrics_loc": "column", "metrics_type": "['Succ. (%)', '#Turns']", "target_entity": "['Ours-Kernel']", "table_html": "
Succ. (%) #Turns
System || Retrieval 9.8 3.26
System || Retrieval-Stgy 67.2 6.56
System || Ours-PMI 47.4 5.12
System || Ours-Neural 51.6 4.29
System || Ours-Kernel 75 4.2
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P19-1565", "page_no": 8, "dir": "acl2019", "description": "Table 4 shows the results of 500 simulations for each of the comparison systems. Our system with Kernel transition obtains the highest success rate, significantly improving over other approaches. The success rate of the base Retrieval agent is lower than 10%, which proves that a chitchat agent without a target-guided strategy can hardly accomplish our task. The Retrieval-Stgy agent has a relatively high success rate, while taking more turns (6.56) to accomplish this. This is partially due to the lack of coarse-grained keyword modeling and transition. We further note that, in the Kernel system, around 81% of predicted keywords eventually occur in the produced utterances, indicating that the predicted keywords have a great impact on the retrieval module.", "class_sentence": "[1, 1, 1, 1, 2, 2]", "sentences": "['Table 4 shows the results of 500 simulations for each of the comparison systems.', 'Our system with Kernel transition obtains the highest success rate, significantly improving over other approaches.', 'The success rate of the base Retrieval agent is lower than 10%, which proves that a chitchat agent without a target-guided strategy can hardly accomplish our task.', 'The Retrieval-Stgy agent has a relatively high success rate, while taking more turns (6.56) to accomplish this.', 'This is partially due to the lack of coarse-grained keyword modeling and transition.', 'We further note that, in the Kernel system, around 81% of predicted keywords eventually occur in the produced utterances, indicating that the predicted keywords have a great impact on the retrieval module.']", "header_mention": "[None, ['Ours-Kernel', 'Succ. (%)'], ['Retrieval', 'Succ. (%)'], ['Retrieval-Stgy', '#Turns'], None, None]", "valid": 1, "image_id": "P19-1565-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Results of Self-Play Evaluation.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Succ. (%)#Turns
System || Retrieval9.83.26
System || Retrieval-Stgy67.26.56
System || Ours-PMI47.45.12
System || Ours-Neural51.64.29
System || Ours-Kernel754.2
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Results of Self-Play Evaluation.}\n\\begin{tabular}{lrr}\n\\toprule\n & Succ. (\\%) & \\#Turns \\\\\n\\midrule\nSystem $\\parallel$ Retrieval & 9.8 & 3.26 \\\\\nSystem $\\parallel$ Retrieval-Stgy & 67.2 & 6.56 \\\\\nSystem $\\parallel$ Ours-PMI & 47.4 & 5.12 \\\\\nSystem $\\parallel$ Ours-Neural & 51.6 & 4.29 \\\\\nSystem $\\parallel$ Ours-Kernel & 75 & 4.2 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: Results of Self-Play Evaluation.\n
\n \n Succ. (%)\n \n #Turns\n
\n System || Retrieval\n \n 9.8\n \n 3.26\n
\n System || Retrieval-Stgy\n \n 67.2\n \n 6.56\n
\n System || Ours-PMI\n \n 47.4\n \n 5.12\n
\n System || Ours-Neural\n \n 51.6\n \n 4.29\n
\n System || Ours-Kernel\n \n 75\n \n 4.2\n
\n" }, { "instance_id": "f41b33cd-b35b-4ce8-9c8c-d9df8f715da6", "table_id_paper": "P19-1570table_7", "caption": "Comparison on the SCWS task. Setting (a) for WordCtx2Sense uses \u03bb = 0.1 for all pairs, and setting (b) uses \u03bb = 10\u22123 for pairs containing same target words and \u03bb = 0.1 for all other pairs. Word2Sense, Word2Vec and Word2GM neglect context and compare target words. a numbers reported from (Mu et al., 2017) whose experimental setup we could replicate.", "row_header_level": 2, "row_headers": [ [ "Method", "WordCtx2Sense (a)" ], [ "Method", "WordCtx2Sense (b)" ], [ "Method", "Word2Sense" ], [ "Method", "Word2vec" ], [ "Method", "(Mu et al., 2017)" ], [ "Method", "(Huang et al., 2012)" ], [ "Method", "(Arora et al., 2018)" ], [ "Method", "Word2GM" ], [ "Method", "MSSG.300D.30K" ], [ "Method", "MSSG.300D.6K" ] ], "column_header_level": 1, "column_headers": [ [ "Pearson-coefficient" ] ], "contents": [ [ "0.666" ], [ "0.67" ], [ "0.644" ], [ "0.651" ], [ "0.637" ], [ "0.657" ], [ "0.652" ], [ "0.655" ], [ "0.679a" ], [ "0.678a" ] ], "metrics_loc": "column", "metrics_type": "['Pearson-coefficient']", "target_entity": "['WordCtx2Sense (a)', 'WordCtx2Sense (b)']", "table_html": "
Pearson-coefficient
Method || WordCtx2Sense (a) 0.666
Method || WordCtx2Sense (b) 0.67
Method || Word2Sense 0.644
Method || Word2vec 0.651
Method || (Mu et al., 2017) 0.637
Method || (Huang et al., 2012) 0.657
Method || (Arora et al., 2018) 0.652
Method || Word2GM 0.655
Method || MSSG.300D.30K 0.679a
Method || MSSG.300D.6K 0.678a
", "table_name": "Table 7", "table_id": "table_7", "paper_id": "P19-1570", "page_no": 9, "dir": "acl2019", "description": "Results . Table 7 shows that sense embeddings using context information perform better than all the existing models, except MSSG models (Neelakantan et al., 2015). Also, computing the embeddings of a word using the contextual information improves results by aprox. 0.025, compared to the case when words embeddings are used directly.", "class_sentence": "[2, 1, 1]", "sentences": "['Results .', 'Table 7 shows that sense embeddings using context information perform better than all the existing models, except MSSG models (Neelakantan et al., 2015).', 'Also, computing the embeddings of a word using the contextual information improves results by aprox. 0.025, compared to the case when words embeddings are used directly.']", "header_mention": "[None, ['WordCtx2Sense (a)', 'WordCtx2Sense (b)', 'MSSG.300D.30K', 'MSSG.300D.6K'], ['WordCtx2Sense (a)', 'WordCtx2Sense (b)']]", "valid": 1, "image_id": "P19-1570-Table7-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Comparison on the SCWS task. Setting (a) for WordCtx2Sense uses \u03bb = 0.1 for all pairs, and setting (b) uses \u03bb = 10\u22123 for pairs containing same target words and \u03bb = 0.1 for all other pairs. Word2Sense, Word2Vec and Word2GM neglect context and compare target words. a numbers reported from (Mu et al., 2017) whose experimental setup we could replicate.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Pearson-coefficient
Method || WordCtx2Sense (a)0.666
Method || WordCtx2Sense (b)0.67
Method || Word2Sense0.644
Method || Word2vec0.651
Method || (Mu et al., 2017)0.637
Method || (Huang et al., 2012)0.657
Method || (Arora et al., 2018)0.652
Method || Word2GM0.655
Method || MSSG.300D.30K0.679a
Method || MSSG.300D.6K0.678a
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{6}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Comparison on the SCWS task. Setting (a) for WordCtx2Sense uses $\\lambda$ $=$ 0.1 for all pairs, and setting (b) uses $\\lambda$ $=$ 10-3 for pairs containing same target words and $\\lambda$ $=$ 0.1 for all other pairs. Word2Sense, Word2Vec and Word2GM neglect context and compare target words. a numbers reported from (Mu et al., 2017) whose experimental setup we could replicate.}\n\\begin{tabular}{ll}\n\\toprule\n & Pearson-coefficient \\\\\n\\midrule\nMethod $\\parallel$ WordCtx2Sense (a) & 0.666 \\\\\nMethod $\\parallel$ WordCtx2Sense (b) & 0.67 \\\\\nMethod $\\parallel$ Word2Sense & 0.644 \\\\\nMethod $\\parallel$ Word2vec & 0.651 \\\\\nMethod $\\parallel$ (Mu et al., 2017) & 0.637 \\\\\nMethod $\\parallel$ (Huang et al., 2012) & 0.657 \\\\\nMethod $\\parallel$ (Arora et al., 2018) & 0.652 \\\\\nMethod $\\parallel$ Word2GM & 0.655 \\\\\nMethod $\\parallel$ MSSG.300D.30K & 0.679a \\\\\nMethod $\\parallel$ MSSG.300D.6K & 0.678a \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 7: Comparison on the SCWS task. Setting (a) for WordCtx2Sense uses \u03bb = 0.1 for all pairs, and setting (b) uses \u03bb = 10\u22123 for pairs containing same target words and \u03bb = 0.1 for all other pairs. Word2Sense, Word2Vec and Word2GM neglect context and compare target words. a numbers reported from (Mu et al., 2017) whose experimental setup we could replicate.\n
\n \n Pearson-coefficient\n
\n Method || WordCtx2Sense (a)\n \n 0.666\n
\n Method || WordCtx2Sense (b)\n \n 0.67\n
\n Method || Word2Sense\n \n 0.644\n
\n Method || Word2vec\n \n 0.651\n
\n Method || (Mu et al., 2017)\n \n 0.637\n
\n Method || (Huang et al., 2012)\n \n 0.657\n
\n Method || (Arora et al., 2018)\n \n 0.652\n
\n Method || Word2GM\n \n 0.655\n
\n Method || MSSG.300D.30K\n \n 0.679a\n
\n Method || MSSG.300D.6K\n \n 0.678a\n
\n" }, { "instance_id": "dfbc3e2e-027b-48cd-adb6-69813e13e61b", "table_id_paper": "P19-1581table_2", "caption": "Medical bilingual lexicon induction results showing the quality of the BWE based dictionaries using 1-best and 5-best translations.", "row_header_level": 2, "row_headers": [ [ "freq", "(Braune et al., 2018)" ], [ "freq", "EU+UFAL+orth" ], [ "rare", "(Braune et al., 2018)" ], [ "rare", "EU+UFAL+orth" ] ], "column_header_level": 1, "column_headers": [ [ "Acc1" ], [ "Acc5" ] ], "contents": [ [ "38.6", "47.4" ], [ "25.9", "40.6" ], [ "26.3", "28.2" ], [ "17.5", "28.8" ] ], "metrics_loc": "column", "metrics_type": "['Acc1', 'Acc5']", "target_entity": "['EU+UFAL+orth']", "table_html": "
Acc1 Acc5
freq || (Braune et al., 2018) 38.6 47.4
freq || EU+UFAL+orth 25.9 40.6
rare || (Braune et al., 2018) 26.3 28.2
rare || EU+UFAL+orth 17.5 28.8
", "table_name": "Table 2", "table_id": "table_2", "paper_id": "P19-1581", "page_no": 4, "dir": "acl2019", "description": "We also compare the quality of our best dictionary (EU+UFAL+orth) to previous work by running bilingual lexicon induction using the test lexicons of Braune et al. (2018) containing frequent and rare medical words respectively. Accuracies of 1-best and 5-best translations in Table 2 show comparable word translation quality to previous work, although we do not employ any task specific steps in contrast to Braune et al. (2018). Note that our dictionary does not contain some of the rare words of the test lexicons which we ignore during evaluation.", "class_sentence": "[2, 1, 2]", "sentences": "['We also compare the quality of our best dictionary (EU+UFAL+orth) to previous work by running bilingual lexicon induction using the test lexicons of Braune et al. (2018) containing frequent and rare medical words respectively.', 'Accuracies of 1-best and 5-best translations in Table 2 show comparable word translation quality to previous work, although we do not employ any task specific steps in contrast to Braune et al. (2018).', 'Note that our dictionary does not contain some of the rare words of the test lexicons which we ignore during evaluation.']", "header_mention": "[['EU+UFAL+orth', '(Braune et al., 2018)', 'freq', 'rare'], ['Acc1', 'Acc5', '(Braune et al., 2018)'], None]", "valid": 1, "image_id": "P19-1581-Table2-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Medical bilingual lexicon induction results showing the quality of the BWE based dictionaries using 1-best and 5-best translations.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Acc1Acc5
freq || (Braune et al., 2018)38.647.4
freq || EU+UFAL+orth25.940.6
rare || (Braune et al., 2018)26.328.2
rare || EU+UFAL+orth17.528.8
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{1}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Medical bilingual lexicon induction results showing the quality of the BWE based dictionaries using 1-best and 5-best translations.}\n\\begin{tabular}{lrr}\n\\toprule\n & Acc1 & Acc5 \\\\\n\\midrule\nfreq $\\parallel$ (Braune et al., 2018) & 38.6 & 47.4 \\\\\nfreq $\\parallel$ EU+UFAL+orth & 25.9 & 40.6 \\\\\nrare $\\parallel$ (Braune et al., 2018) & 26.3 & 28.2 \\\\\nrare $\\parallel$ EU+UFAL+orth & 17.5 & 28.8 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 2: Medical bilingual lexicon induction results showing the quality of the BWE based dictionaries using 1-best and 5-best translations.\n
\n \n Acc1\n \n Acc5\n
\n freq || (Braune et al., 2018)\n \n 38.6\n \n 47.4\n
\n freq || EU+UFAL+orth\n \n 25.9\n \n 40.6\n
\n rare || (Braune et al., 2018)\n \n 26.3\n \n 28.2\n
\n rare || EU+UFAL+orth\n \n 17.5\n \n 28.8\n
\n" }, { "instance_id": "6a9adfbb-c50a-49c6-ba90-b7bdd3fb51c4", "table_id_paper": "P19-1603table_1", "caption": "Automatic evaluation of sentiment analyzers.", "row_header_level": 2, "row_headers": [ [ "Model", "Rule-Based (RB)" ], [ "Model", "Regression Model (RM)" ], [ "Model", "Domain Adversarial (DA)" ] ], "column_header_level": 1, "column_headers": [ [ "H-M SentiCons" ] ], "contents": [ [ "0.936" ], [ "0.846" ], [ "0.747" ] ], "metrics_loc": "column", "metrics_type": "['H-M SentiCons']", "target_entity": "['Rule-Based (RB)']", "table_html": "
H-M SentiCons
Model || Rule-Based (RB) 0.936
Model || Regression Model (RM) 0.846
Model || Domain Adversarial (DA) 0.747
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P19-1603", "page_no": 4, "dir": "acl2019", "description": "Table 1 shows the automatic evaluation results of three sentiment analyzers. We find that: (1) The rule-based method RB performs the best. This accords with the fact that story endings in the ROC-Stories corpus are simple and have relatively obvious emotional words. (2) DA can not improve the performance of sentiment analysis in our task compared to RM. We hypothesize that is because the domains of labeled SST corpus and ROCStories corpus differ too much that affects the performance of domain adaptation.", "class_sentence": "[1, 1, 2, 1, 2]", "sentences": "['Table 1 shows the automatic evaluation results of three sentiment analyzers.', 'We find that: (1) The rule-based method RB performs the best.', 'This accords with the fact that story endings in the ROC-Stories corpus are simple and have relatively obvious emotional words.', '(2) DA can not improve the performance of sentiment analysis in our task compared to RM.', 'We hypothesize that is because the domains of labeled SST corpus and ROCStories corpus differ too much that affects the performance of domain adaptation.']", "header_mention": "[['Model'], ['Rule-Based (RB)'], None, ['Domain Adversarial (DA)', 'Regression Model (RM)'], None]", "valid": 1, "image_id": "P19-1603-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Automatic evaluation of sentiment analyzers.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n H-M SentiCons
Model || Rule-Based (RB)0.936
Model || Regression Model (RM)0.846
Model || Domain Adversarial (DA)0.747
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Automatic evaluation of sentiment analyzers.}\n\\begin{tabular}{lr}\n\\toprule\n & H-M SentiCons \\\\\n\\midrule\nModel $\\parallel$ Rule-Based (RB) & 0.936 \\\\\nModel $\\parallel$ Regression Model (RM) & 0.846 \\\\\nModel $\\parallel$ Domain Adversarial (DA) & 0.747 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Automatic evaluation of sentiment analyzers.\n
\n \n H-M SentiCons\n
\n Model || Rule-Based (RB)\n \n 0.936\n
\n Model || Regression Model (RM)\n \n 0.846\n
\n Model || Domain Adversarial (DA)\n \n 0.747\n
\n" }, { "instance_id": "3ed75f26-fc38-4cc0-b46a-1818017d0388", "table_id_paper": "P19-1613table_4", "caption": "F1 score on the test set of HOTPOTQA distractor and full wiki setting. All numbers from the official leaderboard. All models except BiDAF are concurrent work (not published). DECOMPRC achieves the best result out of models reported to both distractor and full wiki setting.", "row_header_level": 2, "row_headers": [ [ "Model", "DECOMPRC" ], [ "Model", "Cognitive Graph" ], [ "Model", "BERT Plus" ], [ "Model", "MultiQA" ], [ "Model", "DFGN+BERT" ], [ "Model", "QFE" ], [ "Model", "GRN" ], [ "Model", "BiDAF" ] ], "column_header_level": 1, "column_headers": [ [ "Dist F1" ], [ "Open F1" ] ], "contents": [ [ "69.63", "40.65" ], [ "-", "48.87" ], [ "69.76", "-" ], [ "-", "40.23" ], [ "68.49", "-" ], [ "68.06", "38.06" ], [ "66.71", "36.48" ], [ "59.02", "" ] ], "metrics_loc": "column", "metrics_type": "['dist f1', 'open f1']", "target_entity": "['DECOMPRC']", "table_html": "
Dist F1 Open F1
Model || DECOMPRC 69.63 40.65
Model || Cognitive Graph - 48.87
Model || BERT Plus 69.76 -
Model || MultiQA - 40.23
Model || DFGN+BERT 68.49 -
Model || QFE 68.06 38.06
Model || GRN 66.71 36.48
Model || BiDAF 59.02
", "table_name": "Table 4", "table_id": "table_4", "paper_id": "P19-1613", "page_no": 6, "dir": "acl2019", "description": "Finally, Table 4 shows the F1 score on the test set for distractor setting and full wiki setting on the leaderboard. These include unpublished models that are concurrent to our work. DECOMPRC achieves the best result out of models that report both distractor and full wiki setting.", "class_sentence": "[1, 2, 1]", "sentences": "['Finally, Table 4 shows the F1 score on the test set for distractor setting and full wiki setting on the leaderboard.', 'These include unpublished models that are concurrent to our work.', 'DECOMPRC achieves the best result out of models that report both distractor and full wiki setting.']", "header_mention": "[['Dist F1', 'Open F1'], None, ['DECOMPRC', 'Dist F1', 'Open F1']]", "valid": 1, "image_id": "P19-1613-Table4-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

F1 score on the test set of HOTPOTQA distractor and full wiki setting. All numbers from the official leaderboard. All models except BiDAF are concurrent work (not published). DECOMPRC achieves the best result out of models reported to both distractor and full wiki setting.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Dist F1Open F1
Model || DECOMPRC69.6340.65
Model || Cognitive Graph-48.87
Model || BERT Plus69.76-
Model || MultiQA-40.23
Model || DFGN+BERT68.49-
Model || QFE68.0638.06
Model || GRN66.7136.48
Model || BiDAF59.02\n
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{3}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{F1 score on the test set of HOTPOTQA distractor and full wiki setting. All numbers from the official leaderboard. All models except BiDAF are concurrent work (not published). DECOMPRC achieves the best result out of models reported to both distractor and full wiki setting.}\n\\begin{tabular}{lll}\n\\toprule\n & Dist F1 & Open F1 \\\\\n\\midrule\nModel $\\parallel$ DECOMPRC & 69.63 & 40.65 \\\\\nModel $\\parallel$ Cognitive Graph & - & 48.87 \\\\\nModel $\\parallel$ BERT Plus & 69.76 & - \\\\\nModel $\\parallel$ MultiQA & - & 40.23 \\\\\nModel $\\parallel$ DFGN+BERT & 68.49 & - \\\\\nModel $\\parallel$ QFE & 68.06 & 38.06 \\\\\nModel $\\parallel$ GRN & 66.71 & 36.48 \\\\\nModel $\\parallel$ BiDAF & 59.02 & \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 4: F1 score on the test set of HOTPOTQA distractor and full wiki setting. All numbers from the official leaderboard. All models except BiDAF are concurrent work (not published). DECOMPRC achieves the best result out of models reported to both distractor and full wiki setting.\n
\n \n Dist F1\n \n Open F1\n
\n Model || DECOMPRC\n \n 69.63\n \n 40.65\n
\n Model || Cognitive Graph\n \n -\n \n 48.87\n
\n Model || BERT Plus\n \n 69.76\n \n -\n
\n Model || MultiQA\n \n -\n \n 40.23\n
\n Model || DFGN+BERT\n \n 68.49\n \n -\n
\n Model || QFE\n \n 68.06\n \n 38.06\n
\n Model || GRN\n \n 66.71\n \n 36.48\n
\n Model || BiDAF\n \n 59.02\n \n
\n" }, { "instance_id": "5a22401c-5d68-4ff1-a2a3-631332c058d6", "table_id_paper": "P19-1617table_1", "caption": "Performance comparison on the private test set of HotpotQA in the distractor setting. Our DFGN is the second best result on the leaderboard before submission (on March 1st). The baseline model is from Yang et al. (2018) and the results with \u2217 is unpublished. DFGN(Ours)\u2020 refers to the same model with a revised entity graph, whose entities are recognized by a BERT NER model. Note that the result of DFGN(Ours)\u2020 is submitted to the leaderboard during the review process of our paper.", "row_header_level": 2, "row_headers": [ [ "Model", "Baseline Model" ], [ "Model", "GRN\u2217" ], [ "Model", "DFGN(Ours)" ], [ "Model", "QFE\u2217" ], [ "Model", "DFGN(Ours)\u2020" ] ], "column_header_level": 2, "column_headers": [ [ "Answer", "EM" ], [ "Answer", "F1" ], [ "Sup Fact", "EM" ], [ "Sup Fact", "F1" ], [ "Joint", "EM" ], [ "Joint", "F1" ] ], "contents": [ [ "45.60", "59.02", "20.32", "64.49", "10.83", "40.16" ], [ "52.92", "66.71", "52.37", "84.11", "31.77", "58.47" ], [ "55.17", "68.49", "49.85", "81.06", "31.87", "58.23" ], [ "53.86", "68.06", "57.75", "84.49", "34.63", "59.61" ], [ "56.31", "69.69", "51.50", "81.62", "33.62", "59.82" ] ], "metrics_loc": "column", "metrics_type": "['EM', 'F1', 'EM', 'F1', 'EM', 'F1']", "target_entity": "['DFGN(Ours)', 'DFGN(Ours)\u2020']", "table_html": "
Answer || EM Answer || F1 Sup Fact || EM Sup Fact || F1 Joint || EM Joint || F1
Model || Baseline Model 45.60 59.02 20.32 64.49 10.83 40.16
Model || GRN\u2217 52.92 66.71 52.37 84.11 31.77 58.47
Model || DFGN(Ours) 55.17 68.49 49.85 81.06 31.87 58.23
Model || QFE\u2217 53.86 68.06 57.75 84.49 34.63 59.61
Model || DFGN(Ours)\u2020 56.31 69.69 51.50 81.62 33.62 59.82
", "table_name": "Table 1", "table_id": "table_1", "paper_id": "P19-1617", "page_no": 7, "dir": "acl2019", "description": "4.2 Main Results . We first present a comparison between baseline models and our DFGN2. Table 1 shows the performance of different models in the private test set of HotpotQA. From the table we can see that our model achieves the second best result on the leaderboard now3 (on March 1st). Besides, the answer performance and the joint performance of our model are competitive against state-of-the-art unpublished models. We also include the result of our model with a revised entity graph whose entities are recognized by a BERT NER model (Devlin et al., 2018). We fine-tune the pre-trained BERT model on the dataset of the CoNLL\u00e2\u20ac\u212203 NER shared task (Sang and De Meulder, 2003) and use it to extract named entities from the input paragraphs. The results show that our model achieves a 1.5% gain in the joint F1-score with the entity graph built from a better entity recognizer.", "class_sentence": "[0, 2, 1, 1, 1, 1, 2, 1]", "sentences": "['4.2 Main Results .', 'We first present a comparison between baseline models and our DFGN2.', 'Table 1 shows the performance of different models in the private test set of HotpotQA.', 'From the table we can see that our model achieves the second best result on the leaderboard now3 (on March 1st).', 'Besides, the answer performance and the joint performance of our model are competitive against state-of-the-art unpublished models.', 'We also include the result of our model with a revised entity graph whose entities are recognized by a BERT NER model (Devlin et al., 2018).', 'We fine-tune the pre-trained BERT model on the dataset of the CoNLL\u00e2\u20ac\u212203 NER shared task (Sang and De Meulder, 2003) and use it to extract named entities from the input paragraphs.', 'The results show that our model achieves a 1.5% gain in the joint F1-score with the entity graph built from a better entity recognizer.']", "header_mention": "[None, None, None, ['DFGN(Ours)'], ['DFGN(Ours)'], ['DFGN(Ours)\u2020'], ['DFGN(Ours)\u2020'], ['DFGN(Ours)\u2020', 'DFGN(Ours)']]", "valid": 1, "image_id": "P19-1617-Table1-1.png", "venue": "acl", "license": "cc by", "table_xml": "\n \n \n

Performance comparison on the private test set of HotpotQA in the distractor setting. Our DFGN is the second best result on the leaderboard before submission (on March 1st). The baseline model is from Yang et al. (2018) and the results with \u2217 is unpublished. DFGN(Ours)\u2020 refers to the same model with a revised entity graph, whose entities are recognized by a BERT NER model. Note that the result of DFGN(Ours)\u2020 is submitted to the leaderboard during the review process of our paper.

\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Answer || EMAnswer || F1Sup Fact || EMSup Fact || F1Joint || EMJoint || F1
Model || Baseline Model45.6059.0220.3264.4910.8340.16
Model || GRN\u221752.9266.7152.3784.1131.7758.47
Model || DFGN(Ours)55.1768.4949.8581.0631.8758.23
Model || QFE\u221753.8668.0657.7584.4934.6359.61
Model || DFGN(Ours)\u202056.3169.6951.5081.6233.6259.82
\n\n", "table_latex": "\\begin{table}[ht]\n\\centering\n\\setcounter{table}{0}\n\\captionsetup{justification=raggedright, singlelinecheck=false}\n \\caption{Performance comparison on the private test set of HotpotQA in the distractor setting. Our DFGN is the second best result on the leaderboard before submission (on March 1st). The baseline model is from Yang et al. (2018) and the results with $\\ast$ is unpublished. DFGN(Ours)$\\dagger$ refers to the same model with a revised entity graph, whose entities are recognized by a BERT NER model. Note that the result of DFGN(Ours)$\\dagger$ is submitted to the leaderboard during the review process of our paper.}\n\\begin{tabular}{lrrrrrr}\n\\toprule\n & Answer $\\parallel$ EM & Answer $\\parallel$ F1 & Sup Fact $\\parallel$ EM & Sup Fact $\\parallel$ F1 & Joint $\\parallel$ EM & Joint $\\parallel$ F1 \\\\\n\\midrule\nModel $\\parallel$ Baseline Model & 45.6 & 59.02 & 20.32 & 64.49 & 10.83 & 40.16 \\\\\nModel $\\parallel$ GRN$\\ast$ & 52.92 & 66.71 & 52.37 & 84.11 & 31.77 & 58.47 \\\\\nModel $\\parallel$ DFGN(Ours) & 55.17 & 68.49 & 49.85 & 81.06 & 31.87 & 58.23 \\\\\nModel $\\parallel$ QFE$\\ast$ & 53.86 & 68.06 & 57.75 & 84.49 & 34.63 & 59.61 \\\\\nModel $\\parallel$ DFGN(Ours)$\\dagger$ & 56.31 & 69.69 & 51.5 & 81.62 & 33.62 & 59.82 \\\\\n\\bottomrule\n\\end{tabular}\n\n\\end{table}\n", "table_html_clean": "\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n
\n Table 1: Performance comparison on the private test set of HotpotQA in the distractor setting. Our DFGN is the second best result on the leaderboard before submission (on March 1st). The baseline model is from Yang et al. (2018) and the results with \u2217 is unpublished. DFGN(Ours)\u2020 refers to the same model with a revised entity graph, whose entities are recognized by a BERT NER model. Note that the result of DFGN(Ours)\u2020 is submitted to the leaderboard during the review process of our paper.\n
\n \n Answer || EM\n \n Answer || F1\n \n Sup Fact || EM\n \n Sup Fact || F1\n \n Joint || EM\n \n Joint || F1\n
\n Model || Baseline Model\n \n 45.60\n \n 59.02\n \n 20.32\n \n 64.49\n \n 10.83\n \n 40.16\n
\n Model || GRN\u2217\n \n 52.92\n \n 66.71\n \n 52.37\n \n 84.11\n \n 31.77\n \n 58.47\n
\n Model || DFGN(Ours)\n \n 55.17\n \n 68.49\n \n 49.85\n \n 81.06\n \n 31.87\n \n 58.23\n
\n Model || QFE\u2217\n \n 53.86\n \n 68.06\n \n 57.75\n \n 84.49\n \n 34.63\n \n 59.61\n
\n Model || DFGN(Ours)\u2020\n \n 56.31\n \n 69.69\n \n 51.50\n \n 81.62\n \n 33.62\n \n 59.82\n
\n" } ]