Dynamic Context Selection for Document-level Neural Machine Translation via Reinforcement Learning

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 2242–2254, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 2242 Dynamic Context Selection for Document-level Neural Machine Translation via Reinforcement Learning Xiaomian Kang1,2, Yang Zhao1,2, Jiajun Zhang1,2,3, and Chengqing Zong1,2,4 1National Laboratory of Pattern Recognition, Institute of Automation, CAS, Beijing, China 2School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China 3Beijing Academy of Artificial Intelligence, Beijing, China 4CAS Center for Excellence in Brain Science and Intelligence Technology, Beijing, China {xiaomian.kang, yang.zhao, jjzhang, cqzong}@nlpr.ia.ac.cn Abstract Document-level neural machine translation has yielded attractive improvements. How- ever, majority of existing methods roughly use all context sentences in a fixed scope. They neglect the fact that different source sen- tences need different sizes of context. To ad- dress this problem, we propose an effective ap- proach to select dynamic context so that the document-level translation model can utilize the more useful selected context sentences to produce better translations. Specifically, we introduce a selection module that is indepen- dent of the translation module to score each candidate context sentence. Then, we propose two strategies to explicitly select a variable number of context sentences and feed them into the translation module. We train the two modules end-to-end via reinforcement learn- ing. A novel reward is proposed to encour- age the selection and utilization of dynamic context sentences. Experiments demonstrate that our approach can select adaptive con- text sentences for different source sentences, and significantly improves the performance of document-level translation methods. 1 Introduction Although neural machine translation (NMT) has achieved great progress in recent years (Cho et al., 2014; Bahdanau et al., 2015; Luong et al., 2015; Vaswani et al., 2017), when fed an entire docu- ment, standard NMT systems translate sentences in isolation without considering the cross-sentence dependencies. Consequently, document-level neu- ral machine translation (DocNMT) methods are proposed to utilize source-side or target-side inter- sentence contextual information to improve trans- lation quality over sentences in a document (Jean et al., 2017; Wang et al., 2017; Tiedemann and Scherrer, 2017; Tu et al., 2018; Kuang et al., 2018; Junczys-Dowmunt, 2019; Ma et al., 2020). # Test Context Settings Model1 Model2 1 previous 2 sentences 20.84 20.94 2 previous 6 sentences 20.90 21.15 3 select 2 from previous 6 22.03 22.14 4 dynamic size from previous 6 22.90 22.74 Table 1: The BLEU (%) scores with different context settings. “Model1” and “Model2” are trained with pre- vious 2 and 6 context sentences, respectively. Under- lined results indicate that training and test context set- tings are consistent. More recently, researchers of DocNMT mainly focus on exploring various attention-based net- works to leverage the cross-sentence context ef- ficiently, and evaluate the special discourse phe- nomena (Bawden et al., 2018; M¨uller et al., 2018; Voita et al., 2019b; Jwalapuram et al., 2019). How- ever, there is still an issue that has received less attention: which context sentences should be used when translating a source sentence? We conduct an experiment to verify an intu- ition: …

Content truncated to 500 words