The SPECTRANS System Description for the WMT21 Terminology Task

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 813–820 November 10–11, 2021. ©2021 Association for Computational Linguistics 813 The SPECTRANS System Description for the WMT21 Terminology Task Nicolas Ballier1 Dahn Cho1 Bilal Faye1 Zong-You Ke1 Hanna Martikainen2 Mojca Pecman1 Jean-Baptiste Yunès3 Guillaume Wisniewski4 Lichao Zhu4 Maria Zimina-Poirot1 1 CLILLAC-ARP / 3 IRIT / 4 LLF, Université de Paris, F-75013 Paris, France 2 CLESTHIA, Université Sorbonne Nouvelle - Paris 3, F-75005 Paris, France {nicolas.ballier, guillaume.wisniewski,jean-baptiste.yunes, maria.zimina-poirot}@u-paris.fr, mpecman@eila.univ-paris-diderot.fr {dcho0501, biljolefa, zongyou.ke.fr}@gmail.com hanna-julia.martikainen@sorbonne-nouvelle.fr Abstract This paper discusses the WMT 2021 terminol- ogy shared task from a "meta" perspective. We present the results of our experiments using the terminology dataset and the OpenNMT (Klein et al., 2017) and JoeyNMT (Kreutzer et al., 2019) toolkits for the language direction En- glish to French. Our experiment 1 compares the predictions of the two toolkits. Experiment 2 uses OpenNMT to fine-tune the model. We report our results for the task with the evalu- ation script but mostly discuss the linguistic properties of the terminology dataset provided for the task. We provide evidence of the im- portance of text genres across scores, having replicated the evaluation scripts. 1 Introduction In our (traditional) sense, terminological databases are the collection of specialised lexical resources that are generally compiled from corpora, in collab- oration with experts from the field, then analysed and structured according to the type of information recorded in term records: terms, equivalents, defini- tions, synonyms, contexts of use, and related terms (hyperonyms, hyponyms, meronyms, holonyms, etc.). The data thus created are empirical and provide knowledge-based representations of the domain (especially in the case of an ontological approach), underlining conceptual links between terms that can be observed (like meronomy: "X is a part of Y") and potentially represented in concep- tual graphs. For instance, the ARTES database (Pecman and Kübler, 2011), used at Université de Paris in Mas- ters studies for teaching terminology management to translation students (Kübler et al., 2018), adopts such a comprehensive approach to terminology, with specific attention to emerging terminology and complex noun phrases (CNPs) (Kübler et al., 2021). In recent works combining studies on terminology, specialised translation and corpus linguistics, atten- tion has been drawn to CNPs in English which have been demonstrated to cause major difficulties dur- ing translation, both human and machine (Kübler et al., 2021; Maniez, 2017). Moreover, studies have demonstrated an increase of complex compound- ing in specialised texts in English over the last few decades, with, for instance, an overwhelming use of patterns with adjectival and participial com- pound pre-modifiers (e.g. receptor-binding activity, electron-dense aggregates) (Mestivier-Volanschi, 2015). For this WMT21 Terminology workshop, we focused on the linguistic properties of the termi- nological dataset provided. We selected what we believe to be the two best models we produced for the EN-FR track with two different neural toolkits but we mostly took the opportunity to discuss the addition of terminology to neural machine transla- tion. The rest of the paper is organised as follows: section 2 summarises our approaches …

Content truncated to 500 words