A Preliminary Study on the Impact of Lexical Concreteness on Word Sense Disambiguation

A Preliminary Study on the Impact of Lexical Concreteness on Word Senses Disambiguation * Oi Yee Kwong Department of Chinese, Translation and Linguistics, and Language Information Sciences Research Centre City University of Hong Kong Tat Chee Avenue, Kowloon, Hong Kong rlolivia@cityu.edu.hk Abstract. Psychologists have shown that abstract words are harder to understand and often acquired later than concrete words. In this work, we study how the difficulty of automatic word sense disambiguation (WSD) might be affected by this intrinsic property of words, namely the concreteness of a word and its individual senses. We also explore the feasibility of inducing a numerical index for sense and lexical concreteness from dictionary definitions. Analysis of system performance in previous SENSEVAL exercises suggests that concrete words are often easier to disambiguate. The high overall agreement between human ratings and definition-induced ratings is also encouraging. The concreteness factor is worth the attention of computational linguists, particularly in terms of how it bears on the differential information demand of individual words in WSD and how the knowledge of this property could be employed to fine-tune WSD systems to better deal with the lexical sensitivity of the task. Keywords: Lexical concreteness, Word sense disambiguation 1. Introduction The SENSEVAL (and recently SemEval) exercises have revealed a lot of issues on automatic word sense disambiguation (WSD), and allowed researchers to learn more about the linguistic and technical aspects of the task. System performance often depends on many factors, including the feature set, availability of training instances, and language models, amongst others. One important linguistic factor is the fine-grainedness of the sense inventory and the semantic closeness among the senses of a word. To this end, Resnik and Yarowsky (1997) suggested that closely related senses are more difficult for WSD, and therefore systems should be penalised less if they fail to distinguish between similar senses than if they fail to tell distinct senses apart. Despite being a psychologically valid and intrinsic property of words and senses, concreteness is seldom addressed in WSD literature. Psychologists have shown, from lexical decision and naming tasks, that abstract words are harder to understand than concrete words, and are often acquired later (e.g. Bleasdale, 1987; Kroll and Merves, 1986; Yore and Ollila, * The work described in this paper was fully supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. CityU 1508/06H). Copyright 2008 by Oi Yee Kwong 235 22nd Pacific Asia Conference on Language, Information and Computation, pages 235–244 1985). This thus implies differential underlying mechanisms in the representation, development, and processing of word meanings in the mental lexicon. By analogy, the inclusion of the concreteness information in computational lexicons should also benefit natural language processing (NLP) tasks like WSD, in addition to maneuvering only linguistic and technical factors. It should also allow us to study polysemy and sense similarity in a more comprehensive and cognitively plausible way. Hence, we start on a preliminary study on the relationship between concreteness and WSD. In particular, …

Content truncated to 500 words