Hybrid Models for Semantic Classification of Chinese Unknown Words

Proceedings of NAACL HLT 2007, pages 188–195, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Hybrid Models for Semantic Classification of Chinese Unknown Words Xiaofei Lu Department of Linguistics and Applied Language Studies Pennsylvania State University University Park, PA 16802, USA xxl13@psu.edu Abstract This paper addresses the problem of clas- sifying Chinese unknown words into fine-grained semantic categories defined in a Chinese thesaurus. We describe three novel knowledge-based models that capture the relationship between the se- mantic categories of an unknown word and those of its component characters in three different ways. We then combine two of the knowledge-based models with a corpus-based model which classifies unknown words using contextual infor- mation. Experiments show that the knowledge-based models outperform previous methods on the same task, but the use of contextual information does not further improve performance. 1 Introduction Research on semantic annotation has focused primarily on word sense disambiguation (WSD), i.e., the task of determining the appropriate sense for each instance of a polysemous word out of a set of senses defined for the word in some lexi- con. Much less work has been done on semantic classification of unknown words, i.e., words that are not listed in the lexicon. However, real texts typically contain a large number of unknown words. Successful classification of unknown words is not only useful for lexical acquisition, but also necessary for natural language process- ing (NLP) tasks that require semantic annotation. This paper addresses the problem of classify- ing Chinese unknown words into fine-grained semantic categories defined in a Chinese thesau- rus, Cilin (Mei et al., 1984). This thesaurus clas- sifies over 70,000 words into 12 major catego- ries, including human (A), concrete object (B), time and space (C), abstract object (D), attributes (E), actions (F), mental activities (G), activities (H), physical states (I), relations (J), auxiliaries (K), and honorifics (L). The 12 major categories are further divided into 94 medium categories, which in turn are subdivided into 1428 small categories. Each small category contains syno- nyms that are close in meaning. For example, under the major category D, the medium cate- gory Dm groups all words that refer to institu- tions, and the small category Dm05 groups all words that refer to educational institutions, e.g., 学校 xuéxiào ‘school’. Unknown word classifi- cation involves a much larger search space than WSD. In classifying words into small categories in Cilin, the search space for a polysemous known word consists of all the categories the word belongs to, but that for an unknown word consists of all the 1428 small categories. Research on WSD has concentrated on using contextual information, which may be limited for infrequent unknown words. On the other hand, Chinese characters carry semantic infor- mation that is useful for predicting the semantic properties of the words containing them. We pre- sent three novel knowledge-based models that capture the relationship between the semantic categories of an unknown word and those of its component characters in three different ways, and combine two of them …

Content truncated to 500 words