Iterative Constrained Back-Translation for Unsupervised Domain Adaptation of Machine Translation

Proceedings of the 29th International Conference on Computational Linguistics, pages 5054–5065 October 12–17, 2022. 5054 Iterative Constrained Back-Translation for Unsupervised Domain Adaptation of Machine Translation Hongxiao Zhang1, Hui Huang2, Jiale Gao1, Yufeng Chen1∗, Jinan Xu1, Jian Liu1 1Beijing Jiaotong University, Beijing, China 2Harbin Institute of Technology, Harbin, China {hongxiaozhang,jialegao,chenyf,jaxu,jianliu}@bjtu.edu.cn, huanghui_hit@126.com Abstract Back-translation (BT) has been proven to be effective in unsupervised domain adaptation of neural machine translation (NMT). However, the existing back-translation methods mainly improve domain adaptability by generating in-domain pseudo-parallel data that contains sentence-structural knowledge, paying less at- tention to the in-domain lexical knowledge, which may lead to poor translation of unseen in-domain words. In this paper, we propose an Iterative Constrained Back-Translation (ICBT) method to incorporate in-domain lexical knowl- edge on the basis of BT for unsupervised do- main adaptation of NMT. Specifically, we ap- ply lexical constraints into back-translation to generate pseudo-parallel data with in-domain lexical knowledge, and then perform round-trip iterations to incorporate more lexical knowl- edge. Based on this, we further explore sam- pling strategies of constrained words in ICBT to introduce more targeted lexical knowledge, via domain specificity and confidence estima- tion. Experimental results on four domains show that our approach achieves state-of-the- art results, improving the BLEU score by up to 3.08 compared to the strongest baseline, which demonstrates the effectiveness of our approach. The codes and models are pub- licly available at https://github.com/ zzzxiaohong/ICBT. 1 Introduction Neural machine translation (NMT) has made break- throughs in resource-rich domains (Bahdanau et al., 2015; Vaswani et al., 2017), which requires abun- dant in-domain parallel data (Koehn and Knowles, 2017). Unfortunately, there is no enough paral- lel data for many domains, while the monolingual corpus is much easier to obtain. Therefore, unsu- pervised domain adaptation of NMT, which aims to improve in-domain translation through out-of- domain parallel corpus and in-domain monolingual ∗Yufeng Chen is the corresponding author. Input: 现 指明 香港 按揭 证券 有限公司 为 香港 公营 单位 。 Output: specifying the hong kong abundance company limited as a public unit in hong kong. Back-Translation Input: 现 指明 香港 mortgage 证券 有限公司 为 香港 公营 单位 。 Output: specifying the hong kong mortgage securities limited as a public sector in hong kong. Constrained Back-Translation Target sentence Target sentence with constraint 现 指明 香港 按揭 证券 有限 公司 为 香港 公营 单位 。 现 指明 香港 按揭 证券 有限 公司 为 香港 公营 单位 。 现 指明 香港 mortgage 证券 有 限公司 为 香港 公营 单位 。 现 指明 香港 mortgage 证券 有 限公司 为 香港 公营 单位 。 BT CBT specifying the hong kong mortgage securi- ties limited as a public sector in hong kong. specifying the hong kong mortgage securi- ties limited as a public sector in hong kong. Translated source sentence specifying the hong kong abundance com- pany limited as a public unit in hong kong. specifying the hong kong abundance com- pany limited as a public unit in hong kong. Translated source sentence Figure 1: An example of the English-Chinese trans- lation task …

Content truncated to 500 words