Domain-Specific Sense Distributions and Predominant Sense Acquisition

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 419–426, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Domain-Specific Sense Distributions and Predominant Sense Acquisition Rob Koeling & Diana McCarthy & John Carroll Department of Informatics, University of Sussex Brighton BN1 9QH, UK � robk,dianam,johnca � @sussex.ac.uk Abstract Distributions of the senses of words are often highly skewed. This fact is exploited by word sense disambiguation (WSD) sys- tems which back off to the predominant sense of a word when contextual clues are not strong enough. The domain of a doc- ument has a strong influence on the sense distribution of words, but it is not feasi- ble to produce large manually annotated corpora for every domain of interest. In this paper we describe the construction of three sense annotated corpora in different domains for a sample of English words. We apply an existing method for acquir- ing predominant sense information auto- matically from raw text, and for our sam- ple demonstrate that (1) acquiring such information automatically from a mixed- domain corpus is more accurate than de- riving it from SemCor, and (2) acquiring it automatically from text in the same do- main as the target domain performs best by a large margin. We also show that for an all words WSD task this automatic method is best focussed on words that are salient to the domain, and on words with a different acquired predominant sense in that domain compared to that acquired from a balanced corpus. 1 Introduction From analysis of manually sense tagged corpora, Kilgarriff (2004) has demonstrated that distributions of the senses of words are often highly skewed. Most researchers working on word sense disambiguation (WSD) use manually sense tagged data such as Sem- Cor (Miller et al., 1993) to train statistical classi- fiers, but also use the information in SemCor on the overall sense distribution for each word as a back- off model. In WSD, the heuristic of just choosing the most frequent sense of a word is very powerful, es- pecially for words with highly skewed sense distri- butions (Yarowsky and Florian, 2002). Indeed, only 5 out of the 26 systems in the recent SENSEVAL-3 English all words task (Snyder and Palmer, 2004) outperformed the heuristic of choosing the most fre- quent sense as derived from SemCor (which would give 61.5% precision and recall1). Furthermore, sys- tems that did outperform the first sense heuristic did so only by a small margin (the top score being 65% precision and recall). Over a decade ago, Gale et al. (1992) observed the tendency for one sense of a word to prevail in a given discourse. To take advantage of this, a method for automatically determining the “one sense” given a discourse or document is required. Magnini et al. (2002) have shown that information about the do- main of a document is very useful for WSD. This is because many concepts are specific to particular do- mains, and for many …

Content truncated to 500 words