Simple Features for Chinese Word Sense Disambiguation

Simple Features for Chinese Word Sense Disambiguation Hoa Trang Dang, Ching-yi Chia, Martha Palmer, and Fu-Dong Chiou Department of Computer and Information Science University of Pennsylvania � htd,chingyc,mpalmer,chioufd � @unagi.cis.upenn.edu Abstract In this paper we report on our experiments on au- tomatic Word Sense Disambiguation using a max- imum entropy approach for both English and Chi- nese verbs. We compare the difficulty of the sense- tagging tasks in the two languages and investigate the types of contextual features that are useful for each language. Our experimental results suggest that while richer linguistic features are useful for English WSD, they may not be as beneficial for Chi- nese. 1 Introduction Word Sense Disambiguation (WSD) is a central open problem at the lexical level of Natural Lan- guage Processing (NLP). Highly ambiguous words pose continuing problems for NLP applications. They can lead to irrelevant document retrieval in In- formation Retrieval systems, and inaccurate transla- tions in Machine Translation systems (Palmer et al., 2000). For example, the Chinese word (jian4) has many different senses, one of which can be trans- lated into English as “see”, and another as “show”. Correctly sense-tagging the Chinese word in context can prove to be highly beneficial for lexical choice in Chinese-English machine translation. Several efforts have been made to develop au- tomatic WSD systems that can provide accurate sense tagging (Ide and Veronis, 1998), with a cur- rent emphasis on creating manually sense-tagged data for supervised training of statistical WSD sys- tems, as evidenced by SENSEVAL-1 (Kilgarriff and Palmer, 2000) and SENSEVAL-2 (Edmonds and Cotton, 2001). Highly polysemous verbs, which have several distinct but related senses, pose the greatest challenge for these systems (Palmer et al., 2001). Predicate-argument information and selec- tional restrictions are hypothesized to be particu- larly useful for disambiguating verb senses. Maximum entropy models can be used to solve any classification task and have been applied to a wide range of NLP tasks, including sentence bound- ary detection, part-of-speech tagging, and parsing (Ratnaparkhi, 1998). Assigning sense tags to words in context can be viewed as a classification task sim- ilar to part-of-speech tagging, except that a separate set of tags is required for each vocabulary item to be sense-tagged. Under the maximum entropy frame- work (Berger et al., 1996), evidence from different features can be combined with no assumptions of feature independence. The automatic tagger esti- mates the conditional probability that a word has sense � given that it occurs in context � , where is a conjunction of features. The estimated prob-� ability is derived from feature weights which are determined automatically from training data so as to produce a probability distribution that has max- imum entropy, under the constraint that it is con- sistent with observed evidence. With existing tools for learning maximum entropy models, the bulk of our work is in defining the types of features to look for in the data. Our goal is to see if sense-tagging of verbs can be improved by combining linguistic features that capture information …

Content truncated to 500 words