SYSTRAN @ WMT 2021: Terminology Task

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 842–850 November 10–11, 2021. ©2021 Association for Computational Linguistics 842 SYSTRAN @ WMT 2021: Terminology Task MinhQuang Pham†‡, Antoine Senellart†, Dan Berrebbi†, Josep Crego†, Jean Senellart† † SYSTRAN, 5 rue Feydeau, 75005 Paris, France firstname.lastname@systrangroup.com ‡ LISN, Université Paris-Saclay 91405 Orsay, France firstname.lastname@limsi.fr Abstract This paper describes SYSTRAN submissions to the WMT 2021 terminology shared task. We participate in the English-to-French trans- lation direction with a standard Transformer neural machine translation network that we en- hance with the ability to dynamically include terminology constraints, a very common in- dustrial practice. Two state-of-the-art termi- nology insertion methods are evaluated based (i) on the use of placeholders complemented with morphosyntactic annotation and (ii) on the use of target constraints injected in the source stream. Results show the suitability of the presented approaches in the evaluated sce- nario where terminology is used in a system trained on generic data only. 1 Introduction The high quality obtained by out-of-the-box neural machine translation (NMT) systems (Bojar et al., 2016) has boosted the adoption of automatic trans- lation by the industry and invigorated the research and development on domain adaption and integra- tion of technology in human translation workflows. For instance, combination with translation mem- ories (Bulte and Tezcan, 2019; Xu et al., 2020; Pham et al., 2020), terminology handling (Hasler et al., 2018; Dinu et al., 2019; Michon et al., 2020; Bergmanis and Pinnis, 2021), interactive transla- tion (Peris and Casacuberta, 2019), post-editing modelling (Chatterjee et al., 2020) or dynamic adaptation (Farajian et al., 2017) are all different techniques to make machine translation part of real- life localization workflow. Terminology resources with all their sophisti- cation have been the core building bricks and a continuous challenge to acquire in volume (Senel- lart et al., 2003) for rule-based engines. At the other extreme, they have been reduced to corpus or aligned “phrase pairs” (Schwenk et al., 2008) for Statistical Machine Translation approaches, miss- ing most of their intrinsic linguistic properties. In contrast, neural machine translation operates on word and sentence representations in a continuous space so, on the one hand, it has access to deep actual linguistic knowledge (Conneau et al., 2018) and demonstrates a huge ability to generalize. But on the other hand, results are more difficult to inter- pret (Koehn and Knowles, 2017), and subsequently the translation process is far more complicated to control. Therefore, as for several other linguistic annotations, the challenge is how terminological information can be “passed” to the model. From a human perspective, even though presentation and usage of dictionaries have evolved from ontology (as found in paper dictionary) to corpus-based pre- sentation, looking up terms in a dictionary is the ultimate point of reference for validating the cor- rect term for a specific domain and context. Inline with the conditions of the WMT 2021 terminology shared task, we present English-to- French NMT engines built from abundant generic (out-of-domain) training data. We evaluate several methods to enhance translation engines with …

Content truncated to 500 words