Naver Labs Europe’s Participation in the Robustness, Chat, and Biomedical Tasks at WMT 2020

Proceedings of the 5th Conference on Machine Translation (WMT), pages 462–472 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 462 Naver Labs Europe’s Participation in the Robustness, Chat, and Biomedical Tasks at WMT 2020 Alexandre B´erard Vassilina Nikoulina first.last@naverlabs.com Naver Labs Europe Ioan Calapodescu Jerin Philip∗ IIIT Hyderabad Abstract This paper describes Naver Labs Europe’s participation in the Robustness, Chat, and Biomedical Translation tasks at WMT 2020. We propose a bidirectional German ↔ English model that is multi-domain, robust to noise, and which can translate entire documents (or bilingual dialogues) at once. We use the same ensemble of such models as our primary sub- mission to all three tasks and achieve com- petitive results. We also experiment with lan- guage model pre-training techniques and eval- uate their impact on robustness to noise and out-of-domain translation. For German, Span- ish, Italian, and French to English translation in the Biomedical Task, we also submit our recently released multilingual Covid19NMT model. 1 Introduction We participate in three German ↔ English tasks: Robustness, Chat, and Biomedical. Because these tasks allow the use of the same German-English data, we are able to submit a single model to all of them. We use adapter layers (Bapna and Fi- rat, 2019) to specialize this common model on the provided in-domain data, and obtain a single multi- domain model. 1.1 Task description Robustness Task This task is split into two tracks: 1) a zero-shot translation track whose goal is to make NMT models that are robust to unseen domains; 2) a few-shot translation track, where only a few thousand examples of a new domain will be provided as training data, to try to improve translation quality on this particular domain, while maintaining good quality on the other domains. ∗Work done during the author’s internship at Naver Labs Europe Chat Translation Task The goal of this task is to translate bilingual customer dialogues between two participants (one German-speaking “customer” and one English-speaking “agent”) to the language of the other participant. It combines three chal- lenges: document-level translation (of dialogues), domain adaptation, and noise robustness. Note that the data was originally all in English (even the customer side) and human-translated to German. Biomedical Task This task is a typical domain adaptation task, where we have access to large amounts of generic parallel data, and smaller amounts of in-domain data. The provided test sets are at the document level, which may be useful to our document-level approach. While the data is clean, it contains many numbers, named entities, and compound medical terms, which may require some “robustness tricks” to handle properly. Note that this task is very `a propos, considering the current pandemic situation, in which a good- quality biomedical MT model could be very help- ful for translating guidelines, news articles about COVID-19, or social media reactions. So, in ad- dition to submitting our German ↔ English multi- domain model, we also participate in several lan- guage pairs (German, Spanish, Italian, and French to English) with our recently released …

Content truncated to 500 words