Universal Simultaneous Machine Translation with Mixture-of-Experts Wait-k Policy

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7306–7317 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 7306 Universal Simultaneous Machine Translation with Mixture-of-Experts Wait-k Policy 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 {zhangshaolei20z, fengyang}@ict.ac.cn Abstract Simultaneous machine translation (SiMT) gen- erates translation before reading the entire source sentence and hence it has to trade off between translation quality and latency. To fulfill the requirements of different translation quality and latency in practical applications, the previous methods usually need to train multiple SiMT models for different latency levels, resulting in large computational costs. In this paper, we propose a universal SiMT model with Mixture-of-Experts Wait-k Policy to achieve the best translation quality under arbitrary latency with only one trained model. Specifically, our method employs multi-head attention to accomplish the mixture of experts where each head is treated as a wait-k expert with its own waiting words number, and given a test latency and source inputs, the weights of the experts are accordingly adjusted to pro- duce the best translation. Experiments on three datasets show that our method outperforms all the strong baselines under different latency, in- cluding the state-of-the-art adaptive policy. 1 Introduction Simultaneous machine translation (SiMT) (Cho and Esipova, 2016; Gu et al., 2017; Ma et al., 2019; Arivazhagan et al., 2019) begins outputting trans- lation before reading the entire source sentence and hence has a lower latency compared to full- sentence machine translation. In practical appli- cations, SiMT usually has to fulfill the require- ments with different levels of latency. For exam- ple, a live broadcast requires a lower latency to provide smooth translation while a formal confer- ence focuses on translation quality and allows for a slightly higher latency. Therefore, an excellent SiMT model should be able to maintain high trans- lation quality under different latency levels. However, the existing SiMT methods, which usu- ally employ fixed or adaptive policy, cannot achieve ∗Corresponding author: Yang Feng. 1 3 5 7 9 ktest 23 24 25 26 27 28 29 BLEU ktrain = ktest ktrain = 1 ktrain = 3 ktrain = 5 ktrain = 7 ktrain = 9 Figure 1: Performance of wait-k models with differ- ent ktrain v.s. ktest on IWSLT15 En→Vi SiMT task. ktrain and ktest mean the number of source tokens to wait before performing translation during training and testing, respectively. the best translation performance under different la- tency with only one model (Ma et al., 2019, 2020). With fixed policy, e.g., wait-k policy (Ma et al., 2019), the SiMT model has to wait for a fixed number of source words to be fed and then read one source word and output one target word alter- nately. In wait-k policy, the number of words to wait for can be different during training and testing, denoted as ktrain and ktest respectively, and the latency is determined by …

Content truncated to 500 words