Resolving Sense Ambiguity of Korean Nouns Based on Concept Co-occurrence Information

Resolving Sense Ambiguity of Korean Nouns Based on Concept Co-occurrence Information You-Jin Chung Jong-Hyeok Lee Div. of Electrical and Computer Engineering Pohang University of Science and Technology (POSTECH) and Advanced Information Technology Research Center(AlTrc) San 31 Hyoja-dong, Nam-gu, Pohang 790-784, R. of Korea prizer@postech.ac.kr jhlee@postech.ac.kr fax: +82-54-279-5699 fax: +82-54-279-5699 Abstract From the view point of the linguistic ty- pology, Korean and Japanese have many grammatical similarities which enable it to easily construct a sense-tagged Korean corpus through an existing high-quality Japanese-to-Korean machine translation system. The sense-tagged corpus may serve as a knowledge source to extract useful clues for word sense disambigua- tion (WSD). This paper addresses a dis- ambiguation model for Korean nouns, whose execution is based on the concept codes extracted from the sense-tagged corpus and the semantic similarity values over a thesaurus hierarchy. By the help of the automatically constructed sense- tagged corpus, we overcome the knowl- edge acquisition bottleneck. Also, we show that the performance of word sense disambiguation can be improved by com- bining several base classifiers. In an ex- perimental evaluation, the proposed model using a majority voting achieved an average precision of 77.75% with an improvement over the baseline by 15.00%, which is very promising for real world MT systems. 1 Introduction Generally, a Korean homograph may be translated into a different Japanese equivalent depending on which sense is used in a given context. Thus, noun sense disambiguation is essential to the selection of an appropriate Japanese target word in Korean-to- Japanese translation. Much research on word sense disambiguation has revealed that several different types of infor- mation can contribute to the resolution of lexical ambiguity. These include surrounding words (an unordered set of words surrounding a target word), local collocations (a short sequence of words near a target word, taking word order into account), syntactic relations (selectional restrictions), parts of speech, morphological forms, semantic context, etc (McRoy, 1992, Yarowsky, 1992, Ng and Zelle, 1997). To extract such information, various types of knowledge sources have been utilized such as ma- chine-readable dictionaries (MRD), thesauri, and computational lexicons. Since most MRDs and thesauri were created for human use and display inconsistencies, these resources have clear limita- tions. Sense-tagged corpora have been used as the most useful knowledge source for WSD. However, despite the value of sense-tagged corpora, two ma- jor obstacles impede the acquisition of lexical knowledge from corpora: the difficulties of manu- ally sense-tagging a training corpus, and data sparseness (Ide and Veronis, 1998). Manual sense- tagging of a corpus is extremely costly, and at pre- sent, very few sense-tagged corpora are available. In our WSD approach, we construct a sense- tagged corpus automatically by using a method based on similarities between Korean and Japanese. Our disambiguation model is based on the work of Li et al (2000), especially focusing on the practi- cality of the method for application to real world MT systems. We alleviate the data sparseness problem by adopting a concept-based processing and reduce the number of features to a practical …

Content truncated to 500 words