Learning Optimal Policy for Simultaneous Machine Translation via Binary Search

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2318–2333 July 9-14, 2023 ©2023 Association for Computational Linguistics Learning Optimal Policy for Simultaneous Machine Translation via Binary Search Shoutao Guo 1,2, Shaolei Zhang 1,2, Yang Feng 1,2∗ 1Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS) 2 University of Chinese Academy of Sciences, Beijing, China {guoshoutao22z, zhangshaolei20z, fengyang}@ict.ac.cn Abstract Simultaneous machine translation (SiMT) starts to output translation while reading the source sentence and needs a precise policy to decide when to output the generated translation. Therefore, the policy determines the number of source tokens read during the translation of each target token. However, it is difficult to learn a precise translation policy to achieve good latency-quality trade-offs, because there is no golden policy corresponding to parallel sentences as explicit supervision. In this paper, we present a new method for constructing the optimal policy online via binary search. By employing explicit supervision, our approach enables the SiMT model to learn the optimal policy, which can guide the model in complet- ing the translation during inference. Experi- ments on four translation tasks show that our method can exceed strong baselines across all latency scenarios1 1 Introduction Simultaneous machine translation (SiMT) (Gu et al., 2017; Ma et al., 2019; Arivazhagan et al., 2019; Ma et al., 2020; Zhang et al., 2020), which outputs the generated translation before reading the whole source sentence, is applicable to many real- time scenarios, such as live broadcast and real-time subtitles. To achieve the goal of high translation quality and low latency (Zhang and Feng, 2022b), the SiMT model relies on a policy that determines the number of source tokens to read during the translation of each target token. The translation policy plays a pivotal role in de- termining the performance of SiMT, as an impre- cise policy can lead to degraded translation quality or introduce unnecessary delays, resulting in poor translation performance (Zhang and Feng, 2022c). ∗Corresponding author: Yang Feng. 1Code is available at https://github.com/ictnlp/ BS-SiMT Therefore, it is crucial to establish an optimal pol- icy that achieves good latency-quality trade-offs. However, the absence of a golden policy between the source and target makes it challenging for the SiMT model to acquire the explicit supervision re- quired for learning the optimal policy. According to Zhang et al. (2020), the SiMT model will learn better policy if it is trained with external supervi- sion. Consequently, by constructing the optimal policy between the source and target, we can train the SiMT model, which will then generate transla- tions based on the learned policy during inference. However, the existing methods, including fixed policy and adaptive policy, have limitations in learn- ing the optimal policy due to the lack of appropriate explicit supervision. For fixed policy (Dalvi et al., 2018; Ma et al., 2019; Elbayad et al., 2020; Zhang and Feng, 2021b), the model relies on heuristic rules to generate translations. However, these rules may not prompt …

Content truncated to 500 words