Unsupervised Word Sense Disambiguation Using Bilingual Comparable Corpora

Unsupervised Word Sense Disambiguation Using Bilingual Comparable Corpora Hiroyuki Kaji and Yasutsugu Morimoto Central Research Laboratory, Hitachi, Ltd. 1-280 Higashi-Koigakubo, Kokubunji-shi, Tokyo 185-8601, Japan {kaji, y-morimo}@crl.hitachi.co.jp Abstract An unsupervised method for word sense disam- biguation using a bilingual comparable corpus was developed. First, it extracts statistically significant pairs of related words from the corpus of each lan- guage. Then, aligning pairs of related words translingually, it calculates the correlation between the senses of a first-language polysemous word and the words related to the polysemous word, which can be regarded as clues for determining the most suitable sense. Finally, for each instance of the polysemous word, it selects the sense that maxi- mizes the score, i.e., the sum of the correlations between each sense and the clues appearing in the context of the instance. To overcome both the problem of ambiguity in the translingual alignment of pairs of related words and that of disparity of topical coverage between corpora of different lan- guages, an algorithm for calculating the correlation between senses and clues iteratively was devised. An experiment using Wall Street Journal and Ni- hon Keizai Shimbun corpora showed that the new method has promising performance; namely, the applicability and precision of its sense selection are 88.5% and 77.7%, respectively, averaged over 60 test polysemous words. 1 Introduction Word sense disambiguation (WSD) is an “intermedi- ate” task that is necessary for accomplishing most natural language processing tasks, especially machine translation and information retrieval. A variety of WSD methods have been proposed over the last dec- ade; however, such methods are still immature. In response to this situation, we have developed an unsu- pervised WSD method using bilingual comparable corpora. With the growing amount of texts available in elec- tronic form, data-driven or corpus-based WSD has become popular. The knowledge useful for WSD can be learned from corpora. However, supervised learning methods suffer from the high cost of manually tagging the sense onto each instance of a polysemous word in a training corpus. A number of bootstrapping methods have been proposed to reduce the sense- tagging cost (Hearst 1991; Basili 1997). A variety of unsupervised WSD methods, which use a machine- readable dictionary or thesaurus in addition to a corpus, have also been proposed (Yarowsky 1992; Yarowsky 1995; Karov and Edelman 1998). Bilingual parallel corpora, in which the senses of words in the text of one language are indicated by their counterparts in the text of another language, have also been used in order to avoid manually sense-tagging training data (Brown, et al. 1991). Unlike the previous methods using bilingual cor- pora, our method does not require parallel corpora. The availability of large parallel corpora is extremely limited. In contrast, comparable corpora are available in many domains. The comparability required by our method is very weak: any combination of corpora of different languages in the same domain is acceptable as a comparable corpus. Several types of information are useful for WSD (Ide and Veronis 1998). Three major types are the grammatical characteristics …

Content truncated to 500 words