Psycholinguistics, Lexicography, and Word Sense Disambiguation

Copyright 2012 by Oi Yee Kwong 26th Pacific Asia Conference on Language,Information and Computation pages 408–417 Psycholinguistics, Lexicography, and Word Sense Disambiguation Oi Yee Kwong Department of Chinese, Translation and Linguistics City University of Hong Kong Tat Chee Avenue, Kowloon, Hong Kong Olivia.Kwong@cityu.edu.hk Abstract Mainstream word sense disambiguation systems have relied mostly on supervised approaches. Complex interactions have been observed between learning algorithms and knowledge sources, but the factors underlying such phenomena are under- explored. This calls for more qualitative analysis of disambiguation results, possibly from an inter-disciplinary perspective. The current study thus preliminarily explores the relation between sense concreteness and the linguistic means for sense distinction with reference to the context availability model proposed in psycholinguistics and common practice in corpus-based lexicography. It will be shown that to a certain extent the varied usefulness of individual knowledge sources for target words, nouns in particular, may be related to the concreteness of the meanings concerned, which predicts how the sense is distinguished from other senses of the word in the first place. A better understanding of this relation is expected to inform the design of disambiguation systems which could then combine algorithms and knowledge sources in a genuine lexically sensitive way. 1 Introduction Word sense ambiguities tend to escape people’s awareness in everyday communication, except in deliberately biased artificial examples or when context is severely limited, since otherwise we almost effortlessly resolve them using a variety of linguistic and extra-linguistic knowledge. This wide range of information is often rendered as various knowledge sources in automatic word sense disambiguation (WSD) systems, partially modelled with different feature sets. As exemplified in recent SENSEVAL and SEMEVAL evaluation exercises (e.g. Kilgarriff and Rosenzweig, 1999; Edmonds and Cotton, 2001; Mihalcea et al., 2004), state-of-the-art WSD systems are mostly based on supervised approaches. Machine learning algorithms are trained on sense-tagged examples, using a wide range of features extracted from the text approximating a variety of knowledge sources deemed useful for the purpose. Ensembles of different types of classifiers based on different feature sets with some voting scheme often report better performance than individual classifiers alone, though the advantage may just be marginal. While complex interactions between learning algorithms and knowledge sources have been observed (e.g. Mihalcea, 2002; Yarowsky and Florian, 2002), and although factors like sense granularity, availability of training data, part-of-speech (POS), etc. are found to relate to such interactions in one way or another, the nature underlying such interactions, which points to the lexical sensitivity issue of WSD, is still somehow under-explored. In particular, more qualitative analysis is needed for disambiguation results, possibly from an inter- disciplinary perspective, for a better understanding of the issue. In the current study, we make a preliminary effort in this regard, and attempt to analyse disambiguation results with respect to the relation 408 Copyright 2012 by Oi Yee Kwong 26 th Pacific Asia Conference on Language, Information and Computation pages 408-417 between sense concreteness and the means for sense distinction in the first place. To this end, we …

Content truncated to 500 words