Content-Equivalent Translated Parallel News Corpus and Extension of Domain Adaptation for NMT

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 3616–3622 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 3616 Content-Equivalent Translated Parallel News Corpus and Extension of Domain Adaptation for Neural Machine Translation Hideya Mino 1,3 Hideki Tanaka 2 Hitoshi Ito 1 Isao Goto 1 Ichiro Yamada 1 Takenobu Tokunaga 3 1 NHK Science & Technology Research Laboratories 1-10-11 Kinuta, Setagaya-ku, Tokyo 157-8510, Japan 2 NHK Engineering System 1-10-11 Kinuta, Setagaya-ku, Tokyo 157-8510, Japan 3 Tokyo Institute of Technology 2-12-1 Ookayama, Meguro-ku, Tokyo 152-8550, Japan mino.h-gq@nhk.or.jp, tanaka.hideki@nes.or.jp, {itou.h-ce,goto.i-es,yamada.i-hy}@nhk.or.jp, take@c.titech.ac.jp Abstract In this paper, we deal with two problems in Japanese-English machine translation of news articles. The first problem is the quality of parallel corpora. Neural machine translation (NMT) systems suffer degraded performance when trained with noisy data. Because there is no clean Japanese-English parallel data for news articles, we build a novel parallel news corpus consisting of Japanese news articles translated into English in a content-equivalent manner. This is the first content-equivalent Japanese-English news corpus translated specifically for training NMT systems. The second problem involves the domain-adaptation technique. NMT systems suffer degraded performance when trained with mixed data having different features, such as noisy data and clean data. Though existing domain- adaptation methods try to overcome this problem by using tags to distinguish the differences between corpora, it is not sufficient. We thus extend a domain-adaptation method by using multiple tags to train an NMT model effectively with both the clean corpus and existing parallel news corpora with some types of noise. Experimental results show that our corpus increases the translation quality, and that our domain-adaptation method is more effective for learning with multiple types of corpora than existing domain-adaptation methods are. Keywords: Parallel News Corpus, Japanese-English, Machine Translation, Domain Adaptation, Back-Translation 1. Introduction Recently, the number of foreigners visiting Japan has in- creased, and most of them cannot understand Japanese. The information gap between Japanese and these foreign- ers is a significant problem. Translation of Japanese news into English is one way to address this problem. Devel- opment of neural machine translation (NMT) systems re- quires a huge amount of clean parallel data. There is al- ready work on Japanese-English parallel news corpora such as the Jiji Corpus1 or the newspaper corpus of the Yomi- uri Database Service2. Though these parallel corpora are precious resources for developing news-focused NMT sys- tems, they include much noise for NMT because of two main factors. The first is that some information is of- ten omitted or added in the English sentences, as com- pared with the Japanese sentences, because the English news articles are generated for English-speaking readers through news writing, not just translating. The second is that these corpora have alignment errors because they are constructed with automatic sentence alignment. Figure 1 shows an example of data with noise due to both omission and addition. The noisier the training data becomes, the more the translation quality of NMT systems deteriorates. To alleviate …

Content truncated to 500 words