Combining machine readable lexical resources and bilingual corpora for broad word sense disambiguation

COMBINING MACHINE READABLE LEXICAL RESOURCES AND BILINGUAL CORPORA FOR BROAD WORD SENSE DISAMBIGUATION Jason J.S. Chang, Jen-Nan Chen, Huei-Hong Sheng, and Sur-Jin Ker Department of Computer Science National Tsing Hua University Hsinchu 30043, Taiwan, ROC Tel: +886 35 731-069 Fax: +886 35 723-694 e-mail: jschang@cs.nthu.edu.tw Abstract This paper describes a new approach to word sense disambiguation (WSD) based on automatically acquired "word sense division. The semantically related sense entries in a bilingual dictionary are arranged in clusters using a heuristic labeling algorithm to provide a more complete and appropriate sense division for WSD. Multiple translations of senses serve as outside information for automatic tagging of bilingual corpora and acquisition of WSD rules. We describe and implement a WSD method using the English-Chinese bilingual version (LecDOCE) of the Longman Dictionary of Contemporary English (LDOCE). For this purpose, we draw on information about topics and topical sets in the Longman Lexicon of Contemporary English (LLOCE) to represent and disambiguate LecDOCE senses. Example sentences and their translations from LecDOCE are employed as training materials for WSD, while further examples from the Brown corpus are used for testing. Quantitative results of disambiguating 12 words are also presented. 1. Introduction Word sense disambiguation (WSD) has been found useful in many NLP applications, including information retrieval (McRoy 1992; Krovetz and Croft 1992) and machine translation (Brown et al. 1991; Dagan et al. 1991; Dagan and Itai 1994). Recent work on computational linguistics have paid increasing attention to WSD (Lesk 1986; Schütze 1992; Gale et al. 1992; Yarowsky 1992 and 1995; Luk 1995). Given a polysemous word in running text, the task of WSD involves examining contextual information to determine the intended sense from a predefined set. If the set of senses is chosen to be possible translations of the word of interest, the WSD becomes the problem of lexical selection (Dagan et al. 1991; Dagan and Itai 1994) in machine translation (MT) process. WSD methods can be characterized by how word senses are divided and by how WSD knowledge is represented and acquired. Word sense is an abstract concept frequently based on subjective and subtle distinctions in topic, register, dialect, collocation, part-of-speech and valency (McRoy 1992). Researchers have experimented with various knowledge sources for WSD system, including (1) the defining words in every-day dictionaries (Lesk 1986; Cowie et al. 1992; Luk 1995), (2) indicative words in the context of words listed under a thesaurus category acquired from corpus (Yarowsky 1992; Chen and Chang 1994), (3) bilingual corpora or monolingual corpora in the target language (Gale, Church and Yarowsky 1992; Dagan et al. 1991; Dagan \ 115 and Itai 1994), (4) automatic induced clusters with sublexical representation (Schütze 1992), and (5) hand- crafted lexicon containing knowledge from multiple sources (McRoy 1992). This study is motivated by several observations. First, word-based approaches trained on dictionaries or corpora offer limited coverage for unrestricted text. Lesk (1986) described a word-sense disambiguation technique that is based on the number of overlap between words in a dictionary definition and words in the …

Content truncated to 500 words