Simultaneous Neural Machine Translation with Constituent Label Prediction

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 1124–1134 November 10–11, 2021. ©2021 Association for Computational Linguistics 1124 Simultaneous Neural Machine Translation with Constituent Label Prediction Yasumasa Kano1 Katsuhito Sudoh1,2 Satoshi Nakamura1,2 1Nara Institute of Science and Technology (NAIST), Japan 2Center for Advanced Intelligence Project (AIP), RIKEN, Japan {kano.yasumasa.kw4, sudoh, s-nakamura}@is.naist.jp Abstract Simultaneous translation is a task in which translation begins before the speaker has fin- ished speaking, so it is important to decide when to start the translation process. However, deciding whether to read more input words or start to translate is difficult for language pairs with different word orders such as English and Japanese. Motivated by the concept of pre- reordering, we propose a couple of simple de- cision rules using the label of the next con- stituent predicted by incremental constituent label prediction. In experiments on English- to-Japanese simultaneous translation, the pro- posed method outperformed baselines in the quality-latency trade-off. 1 Introduction Simultaneous machine translation is a task in which the machine starts outputting a translation before reading the entire input sentence. This task is more difficult than full-sentence translation because it translates the initial part of a sentence without the context of the latter part. This involves a trade-off between delay and quality of the translation; using a longer context should improve translation quality at the cost of a longer delay, and vice versa. In practice, we should control the latency so that it’s not too large, but we may also need to allow a long latency depending on the situation. Most of the recent simultaneous translation mod- els (Ma et al., 2019; Arivazhagan et al., 2019; Raf- fel et al., 2017; Arivazhagan et al., 2019; Ma et al., 2020b; Dalvi et al., 2018; Gu et al., 2017; Alinejad et al., 2018; Cho and Esipova, 2016; Zheng et al., 2020, 2019; Zhang et al., 2020) are based on neural machine translation (NMT), although earlier stud- ies were based on statistical machine translation (Rangarajan Sridhar et al., 2013; Grissom II et al., 2014; Oda et al., 2014, 2015). In simultaneous NMT, there are two major approaches: those in which a latency hyperparameter is given before the training and those in which it is given at the time of inference. The former approach requires training a model individually for each pre-defined latency setting, while the latter approach uses a single model for different latency conditions. Most human simulta- neous interpreters would not need such long train- ing to slightly adjust latency, while it takes much more time to learn other languages to develop their translation skill. Therefore, the latter approach is closer to the learning process of human simultane- ous interpreters. wait-k (Ma et al., 2019) is a simple simultaneous NMT method of the former approach that waits k tokens before starting to translate. It also has variants within the latter approach called test-time wait-k, in which k is determined at the inference time. wait-k had better performance than test-time wait-k in that study’s experiments. There …

Content truncated to 500 words