Fine-tuned xlm-roberta-base for sentence-level language tagging across 100 languages. The model predicts BIO-style language tags over tokens, which makes it useful for language identification, code-switch detection, and multilingual document analysis.

Model description

Introducing Polyglot Tagger, a new way to classify multi-lingual documents. By training specifically on token classification on individual sentences, the model generalizes well on a variety of languages, while also behaves as a multi-label classifier, and extracts sentences based on its language.

Intended uses & limitations

This model can be treated as a base model for further fine-tuning on specific language identification extraction tasks. Note that as a general language tagging model, it can potentially get confused from shared language families or from short texts. For example, English and German, Spanish and Portuguese, and Russian and Ukrainian.

The model is trained on a sentence with a minimum of four tokens, so it may not accurately classify very short and ambigous statements. Note that this model is experimental and may produce unexpected results compared to generic text classifiers. It is trained on cleaned text, therefore, "messy" text may unexpectedly produce different results.

Note that Romanized versions of any language may only have minor representation in the training set, such as Romanized Russian, and Hindi.

Training and Evaluation Data

A synthetic training row consists of 1-4 individual and mostly independent sentences extracted from various sources. The actual training and evaluation data, as well as coverage is found in DerivedFunction/lang-ner-v2.

This model is a fine-tuned version of xlm-roberta-base on an unknown dataset. It achieves the following results on the evaluation set:

Loss: 0.0320
Precision: 0.9504
Recall: 0.9654
F1: 0.9579
Accuracy: 0.9918

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 5e-05
train_batch_size: 8
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 18
total_train_batch_size: 144
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
num_epochs: 2
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
0.7227	0.0804	2500	0.1120	0.7901	0.8797	0.8325	0.9723
0.5850	0.1609	5000	0.0982	0.8418	0.9035	0.8716	0.9777
0.5535	0.2413	7500	0.0808	0.8588	0.9160	0.8865	0.9808
0.4680	0.3218	10000	0.0714	0.8758	0.9240	0.8992	0.9818
0.4853	0.4022	12500	0.0619	0.8905	0.9327	0.9111	0.9839
0.4519	0.4827	15000	0.0559	0.8954	0.9359	0.9152	0.9849
0.4386	0.5631	17500	0.0530	0.8987	0.9385	0.9181	0.9858
0.3982	0.6436	20000	0.0521	0.9043	0.9419	0.9227	0.9866
0.3936	0.7240	22500	0.0496	0.9102	0.9447	0.9271	0.9862
0.3740	0.8045	25000	0.0461	0.9214	0.9482	0.9346	0.9882
0.3405	0.8849	27500	0.0462	0.9232	0.9504	0.9366	0.9878
0.3473	0.9654	30000	0.0421	0.9225	0.9525	0.9373	0.9888
0.2723	1.0458	32500	0.0448	0.9272	0.9538	0.9403	0.9888
0.2616	1.1262	35000	0.0397	0.9342	0.9566	0.9453	0.9896
0.2794	1.2067	37500	0.0414	0.9329	0.9570	0.9448	0.9896
0.2431	1.2871	40000	0.0379	0.9423	0.9602	0.9512	0.9903
0.2577	1.3676	42500	0.0373	0.9371	0.9594	0.9482	0.9902
0.2628	1.4480	45000	0.0362	0.9369	0.9607	0.9486	0.9903
0.2684	1.5285	47500	0.0355	0.9430	0.9613	0.9520	0.9907
0.2592	1.6089	50000	0.0362	0.9467	0.9623	0.9544	0.9908
0.2150	1.6894	52500	0.0340	0.9441	0.9632	0.9535	0.9912
0.2187	1.7698	55000	0.0336	0.9487	0.9639	0.9563	0.9914
0.2074	1.8503	57500	0.0332	0.9482	0.9642	0.9562	0.9916
0.2390	1.9307	60000	0.0320	0.9504	0.9654	0.9579	0.9918