LISN @ WMT 2021

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 232–242 November 10–11, 2021. ©2021 Association for Computational Linguistics 232 LISN @ WMT 2021 Jitao Xu Univ. Paris-Saclay, & CNRS, LISN Pham Minh Quang Univ. Paris-Saclay, & CNRS, LISN & Systran {firstname.lastname}@limsi.fr Sadaf Abdul Rauf Univ. Paris-Saclay, & CNRS, LISN François Yvon Univ. Paris-Saclay, & CNRS, LISN Abstract This paper describes LISN’s submissions to two shared tasks at WMT’21. For the biomedical translation task, we have devel- oped resource-heavy systems for the English- French language pair, using both out-of- domain and in-domain corpora. The target genre for this task (scientific abstracts) corre- sponds to texts that often have a standardized structure. Our systems attempt to take this structure into account using a hierarchical sys- tem of sentence-level tags. Translation sys- tems were also prepared for the News task for the French-German language pair. The chal- lenge was to perform unsupervised adaptation to the target domain (financial news). For this, we explored the potential of retrieval-based strategies, where sentences that are similar to test instances are used to prime the decoder. 1 Introduction This paper describes LISN’s1 submissions to the translation shared tasks at WMT’21, where we took part in two shared tasks. For the biomedical transla- tion tasks, we have developed resource-heavy sys- tems for the English-French language pair, using a diversity of out-of-domain and in-domain corpora, thus continuing the efforts reported in (Abdul Rauf et al., 2020). Like for previous years shared task, the target genre (scientific abstract) corresponds to texts that often have a standardized structure comprising typical subsections of one to five lines. Standard subsections report the OBJECTIVE, the METHOD, or the RESULTs of the study. Our sys- tems for this year attempt to take this structure into account using sentence-level tags, with the hope to capture some of the document structure and the phraseology of the domain into account. These systems are documented in Section 2. 1LISN [Laboratoire Inderdisciplinaire des Sciences du Numérique] is the new name of the laboratory formerly known as LIMSI. Translation systems were also prepared for the News task for the French-German language pair. The challenge this year was to perform unsuper- vised adaptation to the target domain (financial news), with no further detail regarding the test data. In particular, the organizers did not release any de- velopment data to tune systems. In this setting, we explored the potential of using a retrieval-based strategy, where sentences that are similar to the test instances are used to help the decoding. In this approach, introduced in (Bulte and Tezcan, 2019) and further explored in (Xu et al., 2020; Pham et al., 2020), translation is a two-step process: a retrieval phase, which identifies sentences that re- semble the source test sentence in parallel corpora. These sentences and their translation are then used to prime the decoder: inserting relevant translations examples in the decoder’s context should help to select the right translations, especially for words and terms from the test domain. These systems are described in …

Content truncated to 500 words