biglam/bnl_ground_truth_newspapers_before_1878

Dataset description

33.000 transcribed text lines from historical newspapers (before 1878) along with the cropped images of the original scans

Text line based OCR 19.000 text lines in Antiqua 14.000 text lines in Fraktur Transcribed using double-keying (99.95% accuracy) Public Domain, CC0 (See copyright notice) Best for training an OCR engine

The newspapers used are:

• Le Gratis luxembourgeois (1857-1858)
• Luxemburger Volks-Freund (1869-1876)
• L'Arlequin (1848-1848)
• Courrier du Grand-Duché de Luxembourg (1844-1868)
• L'Avenir (1868-1871)
• Der Wächter an der Sauer (1849-1869)
• Luxemburger Zeitung (1844-1845)
• Luxemburger Zeitung = Journal de Luxembourg (1858-1859)
• Der Volksfreund (1848-1849)
• Cäcilia (1862-1871)
• Kirchlicher Anzeiger für die Diözese Luxemburg (1871-1878)
• L'Indépendance luxembourgeoise (1871-1878)
• Luxemburger Anzeiger (1856)
• L'Union (1860-1871)
• Diekircher Wochenblatt (1837-1848)
• Das Vaterland (1869-1870)
• D'Wäschfra (1868-1878)
• Luxemburger Bauernzeitung (1857)
• Luxemburger Wort (1848-1878)

URL for this dataset

https://data.bnl.lu/data/historical-newspapers/

Dataset format

Two JSONL files (antiqua.jsonl.gz and fraktur.jsonl.gz) with the follwing fields:

• font is either antiqua or fraktur
• img is the filename of the associated image for the text
• text is the handcorrected double-keyed text transcribed from the image

Sample:

{
  "font": "fraktur",
  "img": "fraktur-000011.png",
  "text": "Vidal die Vollmacht für Paris an. Auch"
}

In addition there are two .zip files with the associated images

Dataset modality

Text and associated Images from Scans

Dataset licence

Creative Commons Public Domain Dedication and Certification

size of dataset

500MB-2GB

Contact details for data custodian

opendata@bnl.etat.lu

Loading the dataset

The transcriptions and images are also published as parquet, which is what load_dataset reads by default. The original .jsonl.gz and .zip files above are unchanged and can still be downloaded directly.

from datasets import load_dataset

ds = load_dataset("biglam/bnl_ground_truth_newspapers_before_1878", split="train")
row = ds[0]
row["image"]                                  # PIL line crop
row["text"]                                   # the double-keyed transcription
ds.features["font"].int2str(row["font"])      # 'antiqua' or 'fraktur'

field	type	notes
image	Image	the cropped line scan
text	string	hand-corrected double-keyed transcription
font	ClassLabel	antiqua (19,282) or fraktur (14,025)
file_name	string	original filename, e.g. fraktur-000011.png

33,307 lines in total. Transcriptions run 1-118 characters (mean 42.7), none empty; line crops are 105-1141 px wide and 24-56 px tall.

Because font is a label rather than a footnote, OCR accuracy can be reported per font group instead of as one blended number - antiqua and fraktur fail differently, and a single CER figure hides which one a model struggles with.

Notes

• The newspapers listed above span German, French and Luxembourgish, but the source deposit records no per-line language and none has been inferred here.
• These are pre-cropped single lines, so they do not test finding or ordering text on a page. The BnL's Raw Ground Truth Pack (uncropped blocks with ALTO) covers that.
• The 99.95% accuracy figure is the BnL's, from the double-keying process; it has not been re-measured here.