LIMSI @ WMT 2020

Proceedings of the 5th Conference on Machine Translation (WMT), pages 803–812 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 803 LIMSI @ WMT 2020 Sadaf Abdul Rauf Univ. Paris-Saclay, & CNRS, LIMSI Jos´e Carlos Rosales Univ. Paris-Saclay, & CNRS, LIMSI & Inria Paris {firstname.lastname}@limsi.fr Pham Minh Quang Univ. Paris-Saclay, & CNRS, LIMSI & Systran Paris Franc¸ois Yvon Univ. Paris-Saclay, & CNRS, LIMSI Abstract This paper describes LIMSI’s submissions to the translation shared tasks at WMT’20. This year we have focused our efforts on the biomedical translation task, developing a resource-heavy system for the translation of medical abstracts from English into French, using back-translated texts, terminological re- sources as well as multiple pre-processing pipelines, including pre-trained representa- tions. Systems were also prepared for the ro- bustness task for translating from English into German; for this large-scale task we devel- oped multi-domain, noise-robust, translation systems aim to handle the two test conditions: zero-shot and few-shot domain adaptation. 1 Introduction This paper describes LIMSI’s submissions to the translation shared tasks at WMT’20. This year we have focused our efforts on the biomedical translation task, developing a resource-heavy sys- tem for the translation of medical abstract from English into French, using back-translated texts, terminological resources as well as multiple pre- processing pipelines, including pre-trained repre- sentations. Systems where also prepared for the robustness task for translating from English into German; for this large-scale task we developed multi-domain, noise-robust, translation systems aim to handle the two test conditions: zero-shot and few-shot domain adaptation. Machine translation for the biomedical domain is gaining interest owing to the unequivocal sig- nificance of medical scientific texts. The vast ma- jority of these texts are published in English and Biomedical MT aims to also make them available in multiple languages. This is a rather challenging task, due to the scope of this domain, and the cor- responding large and open vocabulary, including terms and non-lexical forms (for dates, biomedical entities, measures, etc). The quality of the resulting MT output thus varies depending on the amount of biomedical (in-domain) resources available for each target language. We participated in this years WMT’20 biomedi- cal translation evaluation for English to French di- rection. English-French is a reasonably resourced language pair with respect to Biomedical parallel corpora, allowing us to train our Neural Machine Translation (NMT) (Sutskever et al., 2014) with only in-domain corpora and dispense with the pro- cessing of large out-of-domain data that exist for this language pair. Our main focus for this year’s participation was to develop strong baselines by making the best of auxiliary resources: back trans- lation of monolingual data; partial pre-translation of terms; pre-trained multilingual contextual em- beddings and IR retrieved in domain corpora. Two pre-prossessing pipelines, one using the standard Moses tools1 and subword-nmt (Sennrich et al., 2016b) and other using HuggingFace BERT API were developed and compared. All systems are based on the transformer architecture (Vaswani et al., 2017), or and on the related BERT-fused transformer model of Zhu et al. (2020). If our base- …

Content truncated to 500 words