Lingua Custodia’s Participation at the WMT 2022 Word-Level Auto-completion Shared Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1170 - 1175 December 7-8, 2022 ©2022 Association for Computational Linguistics Lingua Custodia’s participation at the WMT 2022 Word-Level Auto-completion shared task Melissa Ailem, Jinghsu Liu, Jean-Gabriel Barthélemy and Raheel Qader Lingua Custodia, France {melissa.ailem,jingshu.liu,j-g.barthelemy,raheel.qader}@linguacustodia.com Abstract This paper presents Lingua Custodia’s submis- sion to the WMT22 shared task on Word Level Auto-completion (WLAC). We consider two di- rections, namely German-English and English- German. The WLAC task in Neural Machine Translation (NMT) consists in predicting a tar- get word given few human typed characters, the source sentence to translate, as well as some translation context. Inspired by recent work in terminology control, we propose to treat the human typed sequence as a constraint to pre- dict the right word starting by the latter. To do so, the source side of the training data is augmented with both the constraints and the translation context. In addition, following new advances in WLAC, we use a joint optimiza- tion strategy taking into account several types of translation context. The automatic as well as human accuracy obtained with the submitted systems show the effectiveness of the proposed method. 1 Introduction Modern advances in Neural Machine Translation (NMT) (Sutskever et al., 2014; Bahdanau et al., 2014; Luong et al., 2015; Vaswani et al., 2017) gave rise to a new era, where the translation quality significantly surpasses previous statistical machine translation (SMT) models (Och and Ney, 2002; Koehn et al., 2003; Koehn, 2010). Although these approaches generate high quality translations, there is still a long way to go towards meeting human quality. In fact, NMT models can still generate several types of grammatical and/or semantic mistakes, which is not tolerated in sce- narios requiring accurate and prompt translations. These scenarios include for instance the transla- tions of legal and financial documents, where mis- takes are not permitted and can be costly. To over- come this issue, several Computer-aided translation (CAT) systems have been proposed (Knowles and Koehn, 2016; Santy et al., 2019) to refine NMT models. CAT tools include for instance Auto- matic Post-edition (Junczys-Dowmunt and Grund- kiewicz, 2017; Correia and Martins, 2019; Lopes et al., 2019), terminology control (Hokamp and Liu, 2017; Post and Vilar, 2018; Dinu et al., 2019; Ailem et al., 2021) and sentence vs word-level auto- completion (Knowles and Koehn, 2016; Zhao et al., 2020; Li et al., 2021). Figure 1: Given the source sentence to translate, the translation contexts, and the human typed characters, the WLAC task aims to predict a target word starting by the human typed sequence. As illustrated, the word to predict is not necessarily consecutive to the left and right contexts. The current shared task is on Word-Level Auto- Completion (WLAC) methods, whose objective, as illustrated in Figure 1, is to predict a target word given a source sentence, a translation context, and at least one human typed character. WLAC is a cen- tral Computer-aided task as it helps human transla- tors generate diverse translations quickly and effec- …

Content truncated to 500 words