An HMM - based Method for Segmenting Japanese Terms and Keywords based on Domain - Speci c Bilingual Corpora

Much e ort has been devoted to Japanese morphological analysis (Matsumoto et al 1996; Nagata 1994; Papageorgiou 1994; Yamamoto & Masuyama 1997). The general performances of many of them are quite good. For the purpose of processing complex terms and keywords, however, they have some problems and limitations. Firstly, most general purpose morphological analysers do not segment complex words consistently 1 . There are a few morphological analysers (e.g. Takeda & Fujisaki 1987) which is specialised for the analysis of Chinese-origined units, but they cannot be easily extended for other units such as Katakana. In addition, in IR-oriented applications which we have in mind, the performances of the systems with pre-de ned dictionaries tend to be much lower than their average performance, because of the high domain-speci city. So the proper method for segmenting Japanese complex lexical units is needed as a part of extracting translation patterns of terms.