A General Framework for Adaptation of Neural Machine Translation to Simultaneous Translation

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 191–200 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 191 A General Framework for Adaptation of Neural Machine Translation to Simultaneous Translation Yun Chen∗1, Liangyou Li2, Xin Jiang2, Xiao Chen2, Qun Liu2 1Shanghai University of Finance and Economics, Shanghai, China 2Huawei Noah’s Ark Lab, Hong Kong, China yunchen@sufe.edu.cn, {liliangyou, jiang.xin, chen.xiao2, qun.liu}@huawei.com Abstract Despite the success of neural machine trans- lation (NMT), simultaneous neural machine translation (SNMT), the task of translating in real time before a full sentence has been ob- served, remains challenging due to the syn- tactic structure difference and simultaneity re- quirements. In this paper, we propose a gen- eral framework for adapting neural machine translation to translate simultaneously. Our framework contains two parts: prefix transla- tion that utilizes a consecutive NMT model to translate source prefixes and a stopping crite- rion that determines when to stop the prefix translation. Experiments on three translation corpora and two language pairs show the effi- cacy of the proposed framework on balancing the quality and latency in adapting NMT to per- form simultaneous translation. 1 Introduction Simultaneous translation (F¨ugen et al., 2007; Oda et al., 2014; Grissom et al., 2014; Niehues et al., 2016; Cho and Esipova, 2016; Gu et al., 2017; Ma et al., 2018), the task of producing a partial transla- tion of a sentence before the whole input sentence ends, is useful in many scenarios including out- bound tourism, international summit and multilat- eral negotiations. Different from the consecutive translation in which translation quality alone mat- ters, simultaneous translation trades off between translation quality and latency. The syntactic struc- ture difference between the source and target lan- guage makes simultaneous translation more chal- lenging. For example, when translating from a verb-final (SOV) language (e.g., Japanese) to a verb-media (SVO) language (e.g., English), the verb appears much later in the source sequence ∗Part of the work was done when Yun is working at Huawei Noah’s Ark Lab. than in the target language. Some premature trans- lations can lead to significant loss in quality (Ma et al., 2018). Recently, a number of researchers have endeav- ored to explore methods for simultaneous transla- tion in the context of NMT (Bahdanau et al., 2015; Vaswani et al., 2017). Some of them propose so- phisticated training frameworks explicitly designed for simultaneous translation (Ma et al., 2018; Ari- vazhagan et al., 2019). These approaches are either memory inefficient during training (Ma et al., 2018) or with hyper-parameters hard to tune (Arivazha- gan et al., 2019). Others utilize a full-sentence base model to perform simultaneous translation by modifications to the encoder and the decoding process. To match the incremental source con- text, they replace the bidirectional encoder with a left-to-right encoder (Cho and Esipova, 2016; Satija and Pineau, 2016; Gu et al., 2017; Aline- jad et al., 2018) or recompute the encoder hidden states (Zheng et …

Content truncated to 500 words