Simultaneous Translation Policies: From Fixed to Adaptive

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2847–2853 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 2847 Simultaneous Translation Policies: From Fixed to Adaptive Baigong Zheng 1 Kaibo Liu 1 Renjie Zheng 1 Mingbo Ma 1 Hairong Liu 1 Liang Huang 1,2 1Baidu Research, Sunnyvale, CA, USA 2Oregon State University, Corvallis, OR, USA {baigongzheng, kaiboliu, renjiezheng, mingboma}@baidu.com {liuhairong, lianghuang}@baidu.com Abstract Adaptive policies are better than fixed poli- cies for simultaneous translation, since they can flexibly balance the tradeoff between trans- lation quality and latency based on the cur- rent context information. But previous meth- ods on obtaining adaptive policies either rely on complicated training process, or underper- form simple fixed policies. We design an al- gorithm to achieve adaptive policies via a sim- ple heuristic composition of a set of fixed poli- cies. Experiments on Chinese→English and German→English show that our adaptive poli- cies can outperform fixed ones by up to 4 BLEU points for the same latency, and more surprisingly, it even surpasses the BLEU score of full-sentence translation in the greedy mode (and very close to beam mode), but with much lower latency. 1 Introduction Simultaneous translation (ST) aims to provide good translation quality while keeping the latency of translation process as low as possible. This is very important for the scenarios that require simultane- ity, such as international summits and negotiations. For this, human interpreters usually start transla- tion before the source sentence ends. However, this makes the translation process much more challeng- ing than the full-sentence translation, because to balance the translation quality and latency, inter- preters need to make decisions on when to continue translation and when to stop temporarily to wait for more source side information, which are diffi- cult, especially for syntactically divergent language pairs, such as German and English. The above decisions can be considered as two actions: READ (wait for a new source word) and WRITE (emit a translated target word) (Gu et al., 2017). Then we only need to decide which action to choose at each step, and the solution can be repre- sented by a policy. Earlier works (Yarmohammadi tā � he yĕxŭ �� probably bù � not yīnggāi �� should duì � for cĭ � this fùzé �� be responsible He probably should not be responsible for this Target Source He should be responsible for this R R R R R R wait-3 wait-2 wait-1 R probably not Figure 1: An adaptive policy (in bold arrows) com- posed of three wait-k policies (k = 1, 2, 3). et al., 2013; Bangalore et al., 2012; F¨ugen et al., 2007; Sridhar et al., 2013; Jaitly et al., 2016) study policies as a part of speech-to-speech ST system, where the policies usually try to separate the source sentence into several chunks that can be translated safely. Recent works focus on obtaining policies for text-to-text ST, which can be generally divided into two categories: fixed and adaptive. Fixed poli- cies (Ma et al., 2019; …

Content truncated to 500 words