A Multilingual Parallel Corpus for Improving Machine Translation on Southeast Asian Languages

A Multilingual Parallel Corpus for Improving Machine Translation on Southeast Asian Languages Hai-Long Trieu, Le-Minh Nguyen {trieulh, nguyenml}@jaist.ac.jp Japan Advanced Institute of Science and Technology, Nomi, Ishikawa, Japan Abstract Current machine translation systems require large bilingual corpora for training data. With large bilingual corpora, phrase-based and neural-based methods can achieve state-of-the-art performance. Nevertheless, such large bilingual corpora are unavailable for most language pairs called low-resource languages, which causes a bottleneck for the development of ma- chine translation on such languages. For Southeast Asian region, there is a large population with more than five hundred millions people and several languages that can be used popularly in the world, but there are few parallel data for such language pairs. In this work, we built a multilingual parallel corpus for several Southeast Asian languages. Wikipedia articles’ titles and inter-language link records were used to extract parallel titles. Parallel articles were col- lected based on the parallel titles. For each article pair, parallel sentences were extracted based on a length-based and word correspondences sentence alignment method. A multilingual par- allel corpus were built with more than 1.1 million parallel sentences of ten language pairs of Indonesian, Malay, Filipino, Vietnamese and the languages paired with English. Experiments were conducted on the Asian Language Treebank corpus and showed the promising perfor- mance. Additionally, the corpus was utilized for the IWSLT 2015 machine translation shared task on English-Vietnamese and achieved a significant improvement with +1.7 BLEU point on phrase-based systems and +4.5 BLEU point on a state-of-the-art neural-based system. The corpus can be used to improve machine translation and enhance the development of machine translation on the low-resource Southeast Asian languages. 1 Introduction Current machine translation (MT) systems require large bilingual corpora for training data. With large bilingual corpora up to millions of parallel sentences, MT systems achieve the state- of-the-art performance on both phrase-based (Bojar et al., 2013) and neural-based (Sennrich et al., 2016a) methods. Such large bilingual corpora are available on several language pairs such as English-German, English-French, Czech-English, Chinese-English. For low-resource language pairs, which are most of languages in the world (Irvine, 2013; Wang et al., 2016), there are only small bilingual corpora available. This causes a bottleneck for MT on such language pairs. In order to overcome the problem, previous works have made efforts in building bilin- gual corpora from webs such as in (Utiyama and Isahara, 2003; Li and Liu, 2008; Cettolo et al., 2012). The parallel corpora can be extracted from comparable data such as Wikipedia ((S¸tef˘anescu and Ion, 2013; Chu et al., 2015). The previous work contributed for building bilingual corpora automatically for several low-resource language pairs. For Southeast Asian Proceedings of MT Summit XVI, vol.1: Research Track Nagoya, Sep. 18-22, 2017 | p. 268 languages, there are few bilingual corpora on the languages although there are a high popula- tion with more than five hundred millions of people, and there are several languages that can be used popularly in the world such as Indonesian (ranked 12), Vietnamese (ranked 17) …

Content truncated to 500 words