Resources for the MaTOS (Machine Translation for Open Science) ANR project

This repository contains publicly released resources created in the context of the MaTOS ANR project.

1. Post-editing data for the NLP domain

The data in [NLP-corpus] is described in the following paper, published at the EAMT Conference 2024:

Rachel Bawden, Ziqian Peng, Maud Bénard, Éric Villemonte de la Clergerie, Raphaël Esamotunu, Mathilde Huguin, Natalie Kübler, Alexandra Mestivier, Mona Michelot, Laurent Romary, Lichao Zhu and François Yvon. 2024. Translate your Own: a Post-Edition Experiment in the NLP domain. In Proceedings of the 25th Annual Conference of the European Association for Machine Translation. Sheffield, UK. European Association for Machine Translation. To appear.

Please cite the following:

@inproceedings{bawden-etal-2024-translate,
    title = "{Translate your Own: a Post-Edition Experiment in the NLP domain}",
    author = "Rachel Bawden and Ziqian Peng and Maud Bénard and Éric Villemonte de la Clergerie and Raphaël Esamotunu and Mathilde Huguin and Natalie Kübler and Alexandra Mestivier and Mona Michelot and Laurent Romary and Lichao Zhu and François Yvon",
    booktitle = "Proceedings of the 25th Annual Conference of the European Association for Machine Translation",
    month = jun,
    year = "2024",
    address = "Sheffield, UK",
    publisher = "European Association for Machine Translation"
}

Resources available:

• Corpus:
  • The initial NLP corpus extracted from HAL (English article titles and abstracts): NLP-corpus/nlp_corpus.jsonl, a subsection of recent articles (2020 onwards): nlp_corpus.2020-present.jsonl, and translations by the 3 MT systems: nlp_corpus.2020-present-{systran,deepl,etranslation}.jsonl
  • NLP keywords and list of venues for the filtering of HAL articles: nlp_keywords_dedupl.txt and nlp_venues.txt
  • Creation scripts found in corpus-creation-scripts/. See corpus-creation-scripts/README.md for details.
  • Postedits from the NLP community and translators: NLP-corpus/postedits-NLP-community.csv and NLP-corpus/postedits-translators.csv, and the user metadata: NLP-corpus/postedits-user-metadata.csv
• Post-editing interface:
  • The post-editing interface is still live and you can still contribute here: https://postedition.anr-matos.fr.
  • Code for the post-editing interface: postediting-interface/tal/

2. Parallel corpus of EN-FR abstracts in the NLP domain

The data in [NLP-abstracts] correspond to approximately 2,000 abstracts of PhD Theses and journal abstracts in the NLP domain in French associated with their English translation (or vice-versa). These texts have been downloaded from public sources, manually curated, and aligned at the sentence level. It is redistributed under the terms of the CC-BY Licence.

This data has been used in the following paper, published at the 2024 edition of the TALN conference:

Ziqian Peng, Rachel Bawden, and François Yvon. 2024. À propos des difficultés de traduire automatiquement de longs documents. In Actes de la 31ème Conférence sur le Traitement Automatique des Langues Naturelles, volume 1 : articles longs et prises de position, pages 2–21, Toulouse, France. ATALA and AFPC.

If you would like to use this data, please use the following citation:

@inproceedings{peng-etal-2024-propos,
    title = "{{\`A}} propos des difficult{\'e}s de traduire automatiquement de longs documents",
    author = "Peng, Ziqian  and
      Bawden, Rachel  and
      Yvon, Fran{\c{c}}ois",
    editor = "Balaguer, Mathieu  and
      Bendahman, Nihed  and
      Ho-dac, Lydia-Mai  and
      Mauclair, Julie  and
      G Moreno, Jose  and
      Pinquier, Julien",
    booktitle = "Actes de la 31{\`e}me Conf{\'e}rence sur le Traitement Automatique des Langues Naturelles, volume 1 : articles longs et prises de position",
    month = "7",
    year = "2024",
    address = "Toulouse, France",
    publisher = "ATALA and AFPC",
    url = "https://aclanthology.org/2024.jeptalnrecital-taln.1",
    pages = "2--21",
    abstract = "Les nouvelles architectures de traduction automatique sont capables de traiter des segments longs et de surpasser la traduction de phrases isol{\'e}es, laissant entrevoir la possibilit{\'e} de traduire des documents complets. Pour y parvenir, il est n{\'e}cessaire de surmonter un certain nombre de difficult{\'e}s li{\'e}es {\`a} la longueur des documents {\`a} traduire. Dans cette {\'e}tude, nous discutons de la traduction des documents sous l{'}angle de l{'}{\'e}valuation, en essayant de r{\'e}pondre {\`a} une question simple: comment mesurer s{'}il existe une d{\'e}gradation des performances de traduction avec la longueur des documents ? Nos analyses, qui {\'e}valuent des syst{\`e}mes encodeur-d{\'e}codeur et un grand mod{\`e}le de langue {\`a} l{'}aune de plusieurs m{\'e}triques sur une t{\^a}che de traduction de documents scientifiques sugg{\`e}rent que traduire les documents longs d{'}un bloc reste un probl{\`e}me difficile.",
    language = "French",
}