Code-Switching for Enhancing NMT with Pre-Specified Translation

Proceedings of NAACL-HLT 2019, pages 449–459 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 449 Code-Switching for Enhancing NMT with Pre-Specified Translation Kai Song1,2, Yue Zhang3, Heng Yu2, Weihua Luo2, Kun Wang1, Min Zhang1 1 Soochow University, Suzhou, China 2 Alibaba DAMO Academy, Hangzhou, China 3 School of Engineering, Westlake University, Hangzhou, China songkai.sk@alibaba-inc.com, zhangyue@wias.org.cn {yuheng.yh,weihua.luowh}@alibaba-inc.com kwang1994@stu.suda.edu.cn, minzhang@suda.edu.cn Abstract Leveraging user-provided translation to con- strain NMT has practical significance. Exist- ing methods can be classified into two main categories, namely the use of placeholder tags for lexicon words and the use of hard con- straints during decoding. Both methods can hurt translation fidelity for various reasons. We investigate a data augmentation method, making code-switched training data by replac- ing source phrases with their target transla- tions. Our method does not change the NMT model or decoding algorithm, allowing the model to learn lexicon translations by copy- ing source-side target words. Extensive exper- iments show that our method achieves consis- tent improvements over existing approaches, improving translation of constrained words without hurting unconstrained words. 1 Introduction One important research question in domain- specific machine translation (Luong and Manning, 2015) is how to impose translation constraints (Crego et al., 2016; Hokamp and Liu, 2017; Post and Vilar, 2018). As shown in Figure 1 (a), the word “breadboard” can be translated into “切面 包板 (a wooden board that is used to cut bread on)” in the food domain, but “电路板 (a con- struction base for prototyping of electronics)” in the electronic domain. To enhance translation quality, a lexicon can be leveraged for domain- specific or user-provided words (Arthur et al., 2016; Hasler et al., 2018). We investigate the method of leveraging pre-specified translation for NMT using such a lexicon. For leveraging pre-specified translation, one ex- isting approach uses placeholder tags to substitute named entities (Crego et al., 2016; Li et al., 2016; Wang et al., 2017b) or rare words (Luong et al., Input: I want a breadboard Output: � �� �� ���� Constrained: � �� �� ��� I ��� breadboard user-provided or domain-specific dictionary: Input: I want a breadboard Code-switched: I want a ��� Output: � �� �� ��� (a) Constrained NMT (b) Constrained NMT: Our Method want a I want a I want a Figure 1: Constrained NMT 2014) on both the source and target sides during training, so that a model can translate such words by learning to translate placeholder tags. For ex- ample, the i-th named entity in the source sentence is replaced with “tagi”, as well as its correspond- ing translation in the target side. Placeholder tags in the output are replaced with pre-specified trans- lation as a post-processing step. One disadvantage of this approach, however, is that the meaning of the original words in the pre-specified translation is not fully retained, which can be harmful to both adequacy and fluency of the output. Another approach (Hokamp and Liu, 2017; Post and Vilar, 2018) imposes pre-specified transla- tion via lexical constraints, making sure such …

Content truncated to 500 words