Multilingual Simultaneous Neural Machine Translation

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4758–4766 August 1–6, 2021. ©2021 Association for Computational Linguistics 4758 Multilingual Simultaneous Neural Machine Translation Philip Arthur†∗ Dongwon K. Ryu‡∗ †Oracle Digital Assistant, Oracle ‡Department of Data Science and AI, Monash University philip.arthur@oracle.com {dongwon.ryu, gholamreza.haffari}@monash.edu Gholamreza Haffari‡ Abstract Simultaneous machine translation (SIMT) in- volves translating source utterances to the tar- get language in real-time before the speaker utterance completes. This paper proposes the multilingual approach to SIMT, where a sin- gle model simultaneously translates between multiple language-pairs. This not only results in more efficiency in terms of the number of models and parameters (hence simpler deploy- ment), but may also lead to higher performing models by capturing commonalities among the languages. We further explore simple and ef- fective multilingual architectures based on two strong recently proposed SIMT models. Our results on translating from two Germanic lan- guages (German, Dutch) and three Romance languages (French, Italian, Romanian) into En- glish show (i) the single multilingual model is on-par or better than individual models, and (ii) multilingual SIMT models trained based on language families are on-par or better than the universal model trained for all languages.1 1 Introduction Simultaneous translation is the task of incremen- tally generating the translation while the source utterance is gradually spoken. It is crucial in multi- national meetings, e.g., in business and politics, where the online simultaneous translation is re- quired for one or multiple language-pairs. Simulta- neous machine translation (SIMT) is an attempt to address the challenges of this translation scenario, i.e., trading off the translation quality and its la- tency (Cho and Esipova, 2016; Arivazhagan et al., 2019, 2020; Firat et al., 2016b). In this paper, we investigate the multilingual SIMT setting, where a single model simultane- ously translates between multiple language-pairs. 1Star (*) marks a shared first authorship between Philip and Dongwon, where both contributed equally. This work was done when Philip was a research fellow at Monash University. This not only results in more efficiency in terms of the number of models and parameters (hence simpler deployment), but may also lead to higher performing models by capturing commonalities among the languages. The multilingual setting has been successful for the standard offline neural ma- chine translation (NMT) and studied extensively (Johnson et al., 2017; Tan et al., 2019; Aharoni et al., 2019). We explore simple and effective multilingual ar- chitectures based on two strong recently proposed SIMT models, i.e. the WAIT-K (Dalvi et al., 2018) and COUPLED POLICY (Arthur et al., 2020). The former waits to read a fixed number of k input tokens; afterward, it writes (generates) an output token for each newly received input token. The latter learns a policy, via an agent, for an adaptive waiting between reading and writing to reduce the translation delay while maintaining the quality. COUPLED POLICY uses the adaptive waiting, generated from offline word alignments. It contin- ues to read the source tokens until the correspond- ing word alignment to the target token appears. Under these …

Content truncated to 500 words