An Information Retrieval Approach to Sense Ranking

Proceedings of NAACL HLT 2007, pages 348–355, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics An Information Retrieval Approach to Sense Ranking Mirella Lapata and Frank Keller School of Informatics, University of Edinburgh 2 Buccleuch Place, Edinburgh EH8 9LW, UK {mlap,keller}@inf.ed.ac.uk Abstract In word sense disambiguation, choosing the most frequent sense for an ambigu- ous word is a powerful heuristic. However, its usefulness is restricted by the availabil- ity of sense-annotated data. In this paper, we propose an information retrieval-based method for sense ranking that does not re- quire annotated data. The method queries an information retrieval engine to estimate the degree of association between a word and its sense descriptions. Experiments on the Senseval test materials yield state-of- the-art performance. We also show that the estimated sense frequencies correlate reli- ably with native speakers’ intuitions. 1 Introduction Word sense disambiguation (WSD), the ability to identify the intended meanings (senses) of words in context, is crucial for accomplishing many NLP tasks that require semantic processing. Examples in- clude paraphrase acquisition, discourse parsing, or metonymy resolution. Applications such as machine translation (Vickrey et al., 2005) and information re- trieval (Stokoe, 2005) have also been shown to ben- efit from WSD. Given the importance of WSD for basic NLP tasks and multilingual applications, much work has focused on the computational treatment of sense ambiguity, primarily using data-driven methods. Most accurate WSD systems to date are super- vised and rely on the availability of training data (see Yarowsky and Florian 2002; Mihalcea and Ed- monds 2004 and the references therein). Although supervised methods typically achieve better perfor- mance than unsupervised alternatives, their appli- cability is limited to those words for which sense labeled data exists, and their accuracy is strongly correlated with the amount of labeled data avail- able. Furthermore, current supervised approaches rarely outperform the simple heuristic of choosing the most common or dominant sense in the train- ing data (henceforth “the first sense heuristic”), de- spite taking local context into account. One reason for this is the highly skewed distribution of word senses (McCarthy et al., 2004a). A large number of frequent content words is often associated with only one dominant sense. Obtaining the first sense via annotation is ob- viously costly and time consuming. Sense anno- tated corpora are not readily available for different languages or indeed sense inventories. Moreover, a word’s dominant sense will vary across domains and text genres (the word court in legal documents will most likely mean tribunal rather than yard). It is therefore not surprising that recent work (Mc- Carthy et al., 2004a; Mohammad and Hirst, 2006; Brody et al., 2006) attempts to alleviate the anno- tation bottleneck by inferring the first sense auto- matically from raw text. Automatically acquired first senses will undoubtedly be noisy when compared to human annotations. Nevertheless, they can be use- fully employed in two important tasks: (a) to create preliminary annotations, thus supporting the “anno- tate automatically, correct manually” methodology used to provide high volume annotation in the …

Content truncated to 500 words