Simultaneous Translation with Flexible Policy via Restricted Imitation Learning

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5816–5822 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 5816 Simultaneous Translation with Flexible Policy via Restricted Imitation Learning Baigong Zheng 1,∗ Renjie Zheng 2,∗ Mingbo Ma 1,∗ Liang Huang 1,2 1Baidu Research, Sunnyvale, CA, USA 2Oregon State University, Corvallis, OR, USA {baigongzheng, mingboma}@baidu.com zrenj11@gmail.com Abstract Simultaneous translation is widely useful but remains one of the most difficult tasks in NLP. Previous work either uses fixed-latency poli- cies, or train a complicated two-staged model using reinforcement learning. We propose a much simpler single model that adds a “delay” token to the target vocabulary, and design a restricted dynamic oracle to greatly simplify training. Experiments on Chinese↔English simultaneous translation show that our work leads to flexible policies that achieve better BLEU scores and lower latencies compared to both fixed and RL-learned policies. 1 Introduction Simultaneous translation, which translates sen- tences before they are finished, is useful in many scenarios such as international conferences, sum- mits, and negotiations. However, it is widely con- sidered one of the most challenging tasks in NLP, and one of the holy grails of AI (Grissom II et al., 2014). A major challenge in simultaneous trans- lation is the word order difference between the source and target languages, e.g., between SOV languages (German, Japanese, etc.) and SVO lan- guages (English, Chinese, etc.). Simultaneous translation is previously studied as a part of real-time speech recognition sys- tem (Yarmohammadi et al., 2013; Bangalore et al., 2012; F¨ugen et al., 2007; Sridhar et al., 2013; Jaitly et al., 2016; Graves et al., 2013). Re- cently, there have been two encouraging efforts in this problem with promising but limited success. Gu et al. (2017) propose a complicated two-stage model that is also trained in two stages. The base model, responsible for producing target words, is a conventional full-sentence seq2seq model, and on top of that, the READ/WRITE (R/W) model decides, at every step, whether to wait for an- other source word (READ) or to emit a target word ∗These authors contributed equally. Chinese 我 得到 有关 方面 的 回应 pinyin wˇo d´ed`ao yˇougu¯an f¯angmi`an de hu´ıy`ıng gloss I receive relevant party ’s response wait-1 policy I received thanks from relevant parties wait-5 policy I received re- sponses from relevant parties adaptive policy I received responses from relevant parties Table 1: A Chinese-to-English translation example. Wait-1 policy makes a mistake on guessing thanks from while wait-5 policy has high latency. The adaptive policy can wait for more information to avoid guesses while maintaining low latency. (WRITE) using the pretrained base model. This R/W model is trained by reinforcement learning (RL) method without updating the base model. Ma et al. (2018), on the other hand, propose a much simpler architecture, which only need one model and can be trained with end-to-end local training method. However, their model follows a fixed-latency policy, which inevitably needs to guess future content during translation. Table 1 gives an example …

Content truncated to 500 words