Translation-based Supervision for Policy Generation in Simultaneous Neural Machine Translation

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1734–1744 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 1734 Translation-based Supervision for Policy Generation in Simultaneous Neural Machine Translation Ashkan Alinejad, Hassan S. Shavarani and Anoop Sarkar School of Computing Science Simon Fraser University BC, Canada {aalineja,sshavara,anoop}@sfu.ca Abstract In simultaneous machine translation, finding an agent with the optimal action sequence of reads and writes that maintain a high level of translation quality while minimizing the aver- age lag in producing target tokens remains an extremely challenging problem. We propose a novel supervised learning approach for train- ing an agent that can detect the minimum num- ber of reads required for generating each target token by comparing simultaneous translations against full-sentence translations during train- ing to generate oracle action sequences. These oracle sequences can then be used to train a su- pervised model for action generation at infer- ence time. Our approach provides an alterna- tive to current heuristic methods in simultane- ous translation by introducing a new training objective, which is easier to train than previous attempts at training the agent using reinforce- ment learning techniques for this task. Our ex- perimental results show that our novel training method for action generation produces much higher quality translations while minimizing the average lag in simultaneous translation. 1 Introduction Simultaneous Machine Translation focuses on the real-time translation of the stream of utterances in the source language to the target language. The essence of simultaneous translation imposes a trade-off between translation quality and the delay in the delivery of the translated utterances. Finding the optimal segments in the input stream is one important task to balance the delay and the trans- lation quality. Segmentation based on the input sentence structure (Ryu et al., 2006; Oda et al., 2014; Shavarani et al., 2015) has been explored in previous work. The other approach to reach the optimal segmen- tation strategy (policy) is to define the segmenta- tion problem in reinforcement learning framework (Satija and Pineau, 2016; Gu et al., 2017; Aline- jad et al., 2018) in which the segmentation agent chooses among two possible actions of READ (waiting to receive more input) and WRITE (pro- ducing an output based on the available input in- ventory). The segmentation agents can either use a fixed- latency policy (Dalvi et al., 2018; Ma et al., 2019a) or use an adaptive policy (Grissom II et al., 2014). In the former, the agent waits for receiving a cer- tain number of input tokens and performs pairs of one WRITE and one READ afterward. In the lat- ter, the agent looks for a specific signal in input to permit a WRITE action; a signal that might come from a parsing feature (Grissom II et al., 2014), a stochastic classifier (Gu et al., 2017), or the atten- tion module of the translation model (Arivazhagan et al., 2019; Ma et al., 2020). Recently, imitation learning has been consid- ered to train adaptive policies. This thread of research focuses on …

Content truncated to 500 words