|
|
--- |
|
|
license: cc-by-4.0 |
|
|
--- |
|
|
|
|
|
# OJ4OCRMT: A Large Multilingual Dataset for OCR-MT Evaluation |
|
|
|
|
|
[Paul McNamee, Kevin Duh, Cameron Carpenter, Ron Colaianni, Nolan King, and Kenton Murray. *Proceedings of Machine Translation Summit XX, Vol. 1: Research Track* June 23-27, 2025, Geneva, Switzerland.](https://aclanthology.org/2025.mtsummit-1.9/) |
|
|
|
|
|
The OJ4OCRMT dataset contains source PDF files, rendered images in |
|
|
three resolutions, and text files (both raw extractions, and |
|
|
sentence-boundary split files). There are two partitions, 'dev' |
|
|
and 'test'. Each contains over 1,000 pages of content, with |
|
|
the PDFs, PNGs, and text files available in 23 EU languages. |
|
|
|
|
|
The dataset is designed to support evaluating systems for translation |
|
|
of document images between any pair of 23 European languages. |
|
|
|
|
|
The dev partition contains 1,656 pages from 2022. Of these 1,412 |
|
|
(85%) are deemed regular; 193 (12%) contain a table; and, 51 (3%) |
|
|
contain a 'figure'. The test partition contains 1,119 pages, all from |
|
|
2023. 979 (87%) are regular; 98 (9%) contain a table; and, 42 (4%) |
|
|
contain a 'figure'. Note, all 2,772 pages have translations in all 23 |
|
|
languages. The languages are: Bulgarian, Croatian, Czech, Danish, |
|
|
Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, |
|
|
Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, |
|
|
Slovak, Slovene, Spanish, and Swedish. See the published paper for |
|
|
details about the dataset. |
|
|
|
|
|
Each page is identified by a document identifier and a page number. |
|
|
For example, OJ:C:2022:240 is a document identifier. And page 22 of |
|
|
that document is available in the dev set. At the time of publication |
|
|
this URL gives access to that page online: |
|
|
https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=OJ:C:2022:240:FULL#page=22 |
|
|
|
|
|
For the page identified above, the released dataset contains: |
|
|
- (a) raw: text obtained from the 'pdftotext' command |
|
|
dev/OJ:C:2022:240/raw/OJ:C:2022:240:FULL.en.p-22.txt |
|
|
- (b) sbd: normalized text obtained by cleaning the raw text and performing sentence splitting |
|
|
dev/OJ:C:2022:240/sbd/OJ:C:2022:240:FULL.en.p-22.txt |
|
|
- (c) png72: a PNG image file at 72 dpi resolution |
|
|
dev/OJ:C:2022:240/png72/OJ:C:2022:240:FULL.en.p-22.png |
|
|
- (d) png150: a PNG image file at 150 dpi resolution |
|
|
dev/OJ:C:2022:240/png150/OJ:C:2022:240:FULL.en.p-22.png |
|
|
- (e) png300: a PNG image file at 300 dpi resolution |
|
|
dev/OJ:C:2022:240/png300/OJ:C:2022:240:FULL.en.p-22.png |
|
|
- (f) pdf: a PDF file obtained from the EUR-Lex online portal and using 'pdfseparate' |
|
|
dev/OJ:C:2022:240/pdf/OJ:C:2022:240:FULL.en.p-22.pdf |
|
|
|
|
|
Splits for the partitions are in TSV format with two-columns: |
|
|
docid [tab] page |
|
|
The eight split files are named {dev,test}.{all,regular,table.figure}.txt |
|
|
|
|
|
|
|
|
*Initial release: 6/16/2025* |
