Learning Adaptive Segmentation Policy for Simultaneous Translation

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 2280–2289, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 2280 Learning Adaptive Segmentation Policy for Simultaneous Translation Ruiqing Zhang, Chuanqiang Zhang, Zhongjun He∗, Hua Wu, Haifeng Wang Baidu Inc. No. 10, Shangdi 10th Street, Beijing, 100085, China {zhangruiqing01, zhangchuanqiang, hezhongjun, wu hua, wanghaifeng}@baidu.com Abstract Balancing accuracy and latency is a great chal- lenge for simultaneous translation. To achieve high accuracy, the model usually needs to wait for more streaming text before translation, which results in increased latency. However, keeping low latency would probably hurt accu- racy. Therefore, it is essential to segment the ASR output into appropriate units for transla- tion. Inspired by human interpreters, we pro- pose a novel adaptive segmentation policy for simultaneous translation. The policy learns to segment the source text by considering pos- sible translations produced by the translation model, maintaining consistency between the segmentation and translation. Experimental re- sults on Chinese-English and German-English translation show that our method achieves a better accuracy-latency trade-off over recently proposed state-of-the-art methods. 1 Introduction In recent years, simultaneous translation has at- tracted increasing interest both in research and in- dustry community. It aims at a real-time trans- lation that demands high translation quality and an as-short-as-possible delay between speech and translation output. A typical simultaneous translation system con- sists of an auto-speech-recognition (ASR) sys- tem that transcribes the source speech into source streaming text, and a machine translation (MT) sys- tem that performs the translation from the source into the target text. However, there is a gap between the output of ASR and the input of MT. The MT system takes sentences as input, while the stream- ing ASR output has no segmentation boundaries. Therefore, exploring a policy to split ASR output into appropriate segments becomes a vital issue for simultaneous translation. If translation starts ∗ Corresponding author. before adequate source content is delivered, the translation quality degrades. However, waiting for too much source text increases latency. The policies of recent work generally falls into two classes: • Fixed Policies are hard policies that follow a pre-defined schedule independent of the con- text. They segment the source text based on a fixed length (Ma et al., 2019; Dalvi et al., 2018). For example, the wait-k method (Ma et al., 2019) first reads k source words, and then generates one target word immediately after each subsequent word is received. Poli- cies of this type are simple and easy to imple- ment. However, they do not consider contex- tual information and usually result in a drop in translation accuracy. • Adaptive Policies learn to do segmentation according to dynamic contextual information. They either use a specific model to chunk the streaming source text (Sridhar et al., 2013; Oda et al., 2014; Cho and Esipova, 2016; Gu et al., 2017; Zheng et al., 2019a, 2020) or jointly learn segmentation and translation in an end-to-end framework (Arivazhagan et al., 2019; Zheng et al., 2019b; Ma et al., 2020). The adaptive …

Content truncated to 500 words