STACL: Simultaneous Translation with Implicit Anticipation and Controllable Latency using Prefix-to-Prefix Framework

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3025–3036 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 3025 STACL: Simultaneous Translation with Implicit Anticipation and Controllable Latency using Prefix-to-Prefix Framework ∗ Mingbo Ma1,3 Liang Huang1,3 Hao Xiong2 Renjie Zheng3 Kaibo Liu1,3 Baigong Zheng1 Chuanqiang Zhang2 Zhongjun He2 Hairong Liu1 Xing Li1 Hua Wu2 Haifeng Wang2 1Baidu Research, Sunnyvale, CA, USA 2Baidu, Inc., Beijing, China 3Oregon State University, Corvallis, OR, USA {mingboma, lianghuang, xionghao05, hezhongjun}@baidu.com Abstract Simultaneous translation, which translates sentences before they are finished, is use- ful in many scenarios but is notoriously dif- ficult due to word-order differences. While the conventional seq-to-seq framework is only suitable for full-sentence translation, we pro- pose a novel prefix-to-prefix framework for si- multaneous translation that implicitly learns to anticipate in a single translation model. Within this framework, we present a very sim- ple yet surprisingly effective “wait-k” policy trained to generate the target sentence concur- rently with the source sentence, but always k words behind. Experiments show our strat- egy achieves low latency and reasonable qual- ity (compared to full-sentence translation) on 4 directions: zh↔en and de↔en. 1 Introduction Simultaneous translation aims to automate simul- taneous interpretation, which translates concur- rently with the source-language speech, with a de- lay of only a few seconds. This additive latency is much more desirable than the multiplicative 2× slowdown in consecutive interpretation. With this appealing property, simultaneous in- terpretation has been widely used in many scenar- ios including multilateral organizations (UN/EU), and international summits (APEC/G-20). How- ever, due to the concurrent comprehension and production in two languages, it is extremely chal- lenging and exhausting for humans: the num- ber of qualified simultaneous interpreters world- wide is very limited, and each can only last for about 15-30 minutes in one turn, whose error rates grow exponentially after just minutes of interpret- ing (Moser-Mercer et al., 1998). Moreover, lim- ∗ M.M. and L.H. contributed equally; L.H. conceived the main ideas (prefix-to-prefix and wait-k) and directed the project, while M.M. led the implementations on RNN and Transformer. See example videos, media reports, code, and data at https://simultrans-demo.github.io/. President Bush met with Putin in Moscow Bùshí �� Bush zǒngtǒng �� President zài � at Mòsīkē ��� Moscow yǔ � with Pǔjīng �� Putin huìwù �� meet prediction read write Source side → Target side → Figure 1: Our wait-k model emits target word yt given source-side prefix x1... xt+k−1, often before seeing the corresponding source word (here k=2, outputing y3=“met” before x7=“hu`ıw`u”). Without anticipation, a 5-word wait is needed (dashed arrows). See also Fig. 2. ited memory forces human interpreters to rou- tinely omit source content (He et al., 2016). There- fore, there is a critical need to develop simultane- ous machine translation techniques to reduce the burden of human interpreters and make it more ac- cessible and affordable. Unfortunately, simultaneous translation is also notoriously difficult for machines, due in large part to the diverging word order between the source …

Content truncated to 500 words