Japanese-to-English patent translation system based on domain-adapted word segmentation and post-ordering

Japanese-to-English Patent Translation System based on Domain-adapted Word Segmentation and Post-ordering Katsuhito Sudoh sudoh.katsuhito@lab.ntt.co.jp NTT Communication Science Laboratories, Seika-cho, Soraku-gun, Kyoto 619-0237, Japan / Graduate School of Informatics, Kyoto University, Yoshida-hommachi, Sakyo-ku, Kyoto 606-8501, Japan Masaaki Nagata nagata.masaaki@lab.ntt.co.jp NTT Communication Science Laboratories, Seika-cho, Soraku-gun, Kyoto 619-0237, Japan Shinsuke Mori mori@ar.media.kyoto-u.ac.jp Tatsuya Kawahara kawahara@i.kyoto-u.ac.jp Academic Center for Computing and Media Studies, Kyoto University, Yoshida-hommachi, Sakyo-ku, Kyoto 606-8501, Japan Abstract This paper presents a Japanese-to-English statistical machine translation system specialized for patent translation. Patents are practically useful technical documents, but their translation needs different efforts from general-purpose translation. There are two important problems in the Japanese-to-English patent translation: long distance reordering and lexical translation of many domain-specific terms. We integrated novel lexical translation of domain-specific terms with a syntax-based post-ordering framework that divides the machine translation problem into lexical translation and reordering explicitly for efficient syntax-based translation. The proposed lexical translation consists of a domain-adapted word segmentation and an unknown word transliteration. Experimental results show our system achieves better translation accuracy in BLEU and TER compared to the baseline methods. 1 Introduction Machine translation (MT) is now widely used for various languages and fields. One typical and important use of MT is assimilation, understanding the content of documents written in foreign languages. Among such documents, patents are very important technical documents written in official languages of the countries to which they are filed. MT of the patents is beneficial for industrial use, such as technical surveys in other countries. One practical problem in the patent MT is its domain dependence. Patents usually include long sentences and many technical terms in various fields, which are distinct difference from general-purpose MT. This work focuses on statistical MT (SMT) for patents, from Japanese to English. It is more difficult than English-to-Japanese in: 1) long distance reordering, and 2) word segmentation and lexical translation of domain-specific terms. The first problem is due to large syntactic differences between Japanese and English. Although the reordering in the English-to-Japanese direction can be solved effectively by very simple heuristics called Head Al-Onaizan & Simard (Eds.) Proceedings of AMTA 2014, vol. 1: MT Researchers Vancouver, BC © The Authors 234 Finalization (Isozaki et al., 2010), the reordering in Japanese-to-English is not so straightfor- ward. The second problem results from the Japanese orthography in which there are no explicit word boundaries. General-purpose word segmenters often fail to segment the domain-specific words and those words are translated incorrectly or remain untranslated. Some domain-specific words cannot be translated as unknown words even if they are segmented correctly, due to limited SMT training data. The first problem has been addressed by a syntax-based approach (Yamada and Knight, 2002; Galley et al., 2004; Zollmann and Venugopal, 2006), while most previous studies did not deal with the second problem. Since the lexical translation also af- fects the reordering based on a lexicalized reordering model and an n-gram language model, considering both problems is important for an overall SMT system. The goal of this work is to improve …

Content truncated to 500 words