Compiling French-Japanese Terminologies from the Web

Compiling French-Japanese Terminologies from the Web Xavier Robitaille†, Yasuhiro Sasaki†, Masatsugu Tonoike†, Satoshi Sato‡ and Takehito Utsuro† †Graduate School of Informatics, Kyoto University Yoshida-Honmachi, Sakyo-ku, Kyoto 606-8501 Japan ‡Graduate School of Engineering, Nagoya University Furo-cho, Chikusa-ku, Nagoya 464-8603 Japan {xavier, sasaki, tonoike, utsuro}@pine.kuee.kyoto-u.ac.jp, ssato@nuee.nagoya-u.ac.jp Abstract We propose a method for compiling bi- lingual terminologies of multi-word terms (MWTs) for given translation pairs of seed terms. Traditional methods for bi- lingual terminology compilation exploit parallel texts, while the more recent ones have focused on comparable corpora. We use bilingual corpora collected from the web and tailor made for the seed terms. For each language, we extract from the corpus a set of MWTs pertaining to the seed’s semantic domain, and use a com- positional method to align MWTs from both sets. We increase the coverage of our system by using thesauri and by ap- plying a bootstrap method. Experimental results show high precision and indicate promising prospects for future develop- ments. 1 Introduction Bilingual terminologies have been the center of much interest in computational linguistics. Their applications in machine translation have proven quite effective, and this has fuelled research aim- ing at automating terminology compilation. Early developments focused on their extraction from parallel corpora (Daille et al. (1994), Fung (1995)), which works well but is limited by the scarcity of such resources. Recently, the focus has changed to utilizing comparable corpora, which are easier to obtain in many domains. Most of the proposed methods use the fact that words have comparable contexts across lan- guages. Fung (1998) and Rapp (1999) use so called context vector methods to extract transla- tions of general words. Chiao and Zweigenbaum (2002) and Déjean and Gaussier (2002) apply similar methods to technical domains. Daille and Morin (2005) use specialized comparable cor- pora to extract translations of multi-word terms (MWTs). These methods output a few thousand terms and yield a precision of more or less 80% on the first 10-20 candidates. We argue for the need for systems that output fewer terms, but with a higher precision. Moreover, all the above were conducted on language pairs including English. It would be possible, albeit more difficult, to ob- tain comparable corpora for pairs such as French-Japanese. We will try to remove the need to gather corpora beforehand altogether. To achieve this, we use the web as our only source of data. This idea is not new, and has already been tried by Cao and Li (2002) for base noun phrase translation. They use a compositional method to generate a set of translation candidates from which they select the most likely translation by using empirical evidence from the web. The method we propose takes a translation pair of seed terms in input. First, we collect MWTs semantically similar to the seed in each language. Then, we work out the alignments be- tween the MWTs in both sets. Our intuition is that both seeds have the same related terms across languages, and we believe that this will simplify the alignment …

Content truncated to 500 words