MaTOS: Traduction automatique pour la science ouverte

Maud Bénard, Alexandra Mestivier, Natalie Kübler, Lichao Zhu, Rachel Bawden, Éric Villemonte de la Clergerie, Laurent Romary, Mathilde Huguin, Jean-François Nominé, Ziqian Peng, François YvonARTS@CORIA-TALN 2023 · OpenReview

Abstract

Cette contribution présente le projet MaTOS (Machine Translation for Open Science), qui vise à développer de nouvelles méthodes pour la traduction automatique (TA) intégrale de documents scientifiques entre le français et l’anglais, ainsi que des métriques automatiques pour évaluer la qualité des traductions produites. Pour ce faire, MaTOS s’intéresse (a) au recueil de ressources ouvertes pour la TA spécialisée; (b) à la description des marqueurs de cohérence textuelle pour les articles scientifiques; (c) au développement de nouvelles méthodes de traitement multilingue pour les documents; (d) aux métriques mesurant les progrès de la traduction de documents complets.

Introduction: The Challenge of Scientific Translation

While Neural Machine Translation (NMT) has achieved remarkable success in general domains, scientific literature presents unique challenges that traditional sentence-by-sentence models struggle to overcome. The MaTOS (Machine Translation for Open Science) project addresses these limitations by shifting the focus toward document-level translation. This approach is essential for maintaining terminological consistency and rhetorical coherence across complex research papers, particularly between French and English.

Methodological Innovations: Beyond the Sentence

Current NMT systems often suffer from a "contextual disconnect." When a sentence is translated in isolation, the system may fail to resolve pronouns (anaphora) correctly or may use inconsistent terminology for the same concept throughout a document. MaTOS explores advanced architectures to model the probability of an entire target document given the source.

To manage the computational complexity of long documents, the project utilizes efficient attention mechanisms such as those found in Longformer or BigBird architectures. Furthermore, the team is developing specialized tools to handle terminological variation, ensuring that translations respect the nuances of scientific discourse, such as the alternating use of full terms and their corresponding abbreviations.

Data Mining and Open Science Integration

A significant hurdle in specialized MT is the scarcity of high-quality parallel corpora outside of the biomedical field. MaTOS addresses this by mining massive scientific archives to create new resources.

The HAL Open Science platform serves as a critical resource for data mining and community-based evaluation.

By applying the "SciPar" method to repositories like ISTEX and the HAL open science archive, the project extracts parallel data to train models in underserved domains like Earth Sciences and Natural Language Processing. This effort aligns with the "Helsinki Initiative," promoting multilingualism in science to reduce linguistic biases and make research more accessible to a global audience, including students, journalists, and policy-makers.

Targeted Domains and Terminology

The project focuses on four key areas to ensure technical precision:

  1. Biomedical Sciences: Utilizing established resources like PubMed and WMT-Biomedical challenges.
  2. Natural Language Processing (TAL): Leveraging the ACL60-60 initiative and conference presentation translations.
  3. Earth Sciences: Developing new resources for this historically underserved domain.
  4. Terminology Bases: Integrating specialized bilingual dictionaries and databases like ARTES and Loterre (from INIST).

Evaluation and Human-in-the-Loop

Traditional metrics like BLEU are often too simplistic to capture the nuances of document-level coherence. MaTOS proposes "contrastive evaluation" to specifically check if systems correctly handle long-range linguistic dependencies, such as choosing the correct pronoun based on a noun several sentences prior.

Additionally, the project fosters a participatory approach, encouraging researchers using the HAL archive to post-edit machine-translated abstracts of their own work. This "human-in-the-loop" strategy not only improves translation quality but also generates high-quality parallel data for future model training, creating a virtuous cycle for open science.

Conclusion and Future Directions

MaTOS represents a significant step toward truly automated, document-wide scientific translation. While computational complexity remains a hurdle for encoding full 10-page papers into single vectors, the project's roadmap focuses on direct document-to-document implementation. By combining advanced modeling with a commitment to Open Science principles, MaTOS aims to ensure that scientific knowledge remains a shared global resource, regardless of the original language of publication.

Summary updated

Abstract

Cette contribution présente le projet MaTOS (Machine Translation for Open Science), qui vise à développer de nouvelles méthodes pour la traduction automatique (TA) intégrale de documents scientifiques entre le français et l’anglais, ainsi que des métriques automatiques pour évaluer la qualité des traductions produites. Pour ce faire, MaTOS s’intéresse (a) au recueil de ressources ouvertes pour la TA spécialisée; (b) à la description des marqueurs de cohérence textuelle pour les articles scientifiques; (c) au développement de nouvelles méthodes de traitement multilingue pour les documents; (d) aux métriques mesurant les progrès de la traduction de documents complets.
BibTeX
@inproceedings{benard2023matos,
  author = {Maud Bénard and Alexandra Mestivier and Natalie Kübler and Lichao Zhu and Rachel Bawden and Éric Villemonte de la Clergerie and Laurent Romary and Mathilde Huguin and Jean-François Nominé and Ziqian Peng and François Yvon},
  title = {{MaTOS: Traduction automatique pour la science ouverte}},
  booktitle = {ARTS@CORIA-TALN},
  year = {2023},
  url = {https://openreview.net/forum?id=RfLrX6Uhh1},
}
Figure