CITYU-HIF: WSD with Human-Informed Feature Preference

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 109–112, Prague, June 2007. c⃝2007 Association for Computational Linguistics CITYU-HIF: WSD with Human-Informed Feature Preference Oi Yee Kwong Language Information Sciences Research Centre City University of Hong Kong Tat Chee Avenue, Kowloon, Hong Kong rlolivia@cityu.edu.hk Abstract This paper describes our word sense dis- ambiguation (WSD) system participating in the SemEval-2007 tasks. The core system is a fully supervised system based on a Na- ïve Bayes classifier using multiple knowl- edge sources. Toward a larger goal of in- corporating the intrinsic nature of individ- ual target words in disambiguation, thus in- troducing a cognitive element in automatic WSD, we tried to fine-tune the results ob- tained from the core system with human- informed feature preference, and compared it with automatic feature selection as com- monly practised in statistical WSD. De- spite the insignificant improvement ob- served in this preliminary attempt, more systematic analysis remains to be done for a cognitively plausible account of the fac- tors underlying the lexical sensitivity of WSD, which would inform and enhance the development of WSD systems in return. 1 Introduction In recent years, many research teams all over the world have gained rich experience on word sense disambiguation (WSD) from the shared tasks of the SENSEVAL workshops. The need for multiple knowledge sources has become a golden rule, and the “lexical sensitivity” once remarked by Resnik and Yarowsky (1997) is addressed by various means in statistical classifiers, such as learning an optimal combination of the various knowledge sources for individual target words (e.g. Mihalcea, 2002; Escudero et al., 2004). Another common practice is to use an ensemble of classifiers. As pointed out by Mihalcea et al. (2004), among the participating systems in the SENSEVAL-3 English lexical sample task, “several of the top perform- ance systems are based on combination of multiple classifiers, which shows once again that voting scheme that combine several learning algorithms outperform the accuracy of individual classifiers”. However, the advancement in WSD is rarely ac- companied by any extensive account on the cogni- tive aspects of the task or qualitative analysis of the relation between the disambiguation results and the nature of individual target words underlying the apparent lexical sensitivity of the task. Given that humans apparently use different strategies in making sense of words, it might be beneficial to have such cognitive aspects, including the type and strength of various kinds of semantic association, realised in NLP systems explicitly. Thus in addition to an optimal combination of clas- sifiers alone, to better understand the contribution of different information types for different types of target words, it is important to look at WSD in re- lation to the very intrinsic nature of individual tar- get words, which could comprise many factors such as frequency, abstractness, sense relatedness and parts-of-speech (POS). We thus use the con- cept Information Susceptibility (Kwong, 2005) to refer to the relationship between the intrinsic fea- tures of a target word and its senses, and the effec- tiveness …

Content truncated to 500 words