An Evaluation of Graded Sense Disambiguation using Word Sense Induction

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 189–198, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics An Evaluation of Graded Sense Disambiguation using Word Sense Induction David Jurgens1,2 1HRL Laboratories, LLC Malibu, California, USA 2Department of Computer Science University of California, Los Angeles jurgens@cs.ucla.edu Abstract Word Sense Disambiguation aims to label the sense of a word that best applies in a given context. Graded word sense disambiguation relaxes the single label assumption, allow- ing for multiple sense labels with varying de- grees of applicability. Training multi-label classifiers for such a task requires substan- tial amounts of annotated data, which is cur- rently not available. We consider an alter- nate method of annotating graded senses using Word Sense Induction, which automatically learns the senses and their features from cor- pus properties. Our work proposes three ob- jective to evaluate performance on the graded sense annotation task, and two new methods for mapping between sense inventories using parallel graded sense annotations. We demon- strate that sense induction offers significant promise for accurate graded sense annotation. 1 Introduction Word Sense Disambiguation (WSD) aims to identify the sense of a word in a given context, using a pre- defined sense inventory containing the word’s differ- ent meanings (Navigli, 2009). Traditionally, WSD approaches have assumed that each occurrence of a word is best labeled with a single sense. How- ever, human annotators often disagree about which sense is present (Passonneau et al., 2010), espe- cially in cases where some of the possible senses are closely related (Chugur et al., 2002; McCarthy, 2006; Palmer et al., 2007). Recently, Erk et al. (2009) have shown that in cases of sense ambiguity, a graded notion of sense labeling may be most appropriate and help reduce the ambiguity. Specifically, within a given context, multiple senses of a word may be salient to the reader, with different levels of applicability. For ex- ample, in the sentence • The athlete won the gold metal due to her hard work and dedication. multiple senses could be considered applicable for “won” according to the WordNet 3.0 sense inventory (Fellbaum, 1998): 1. win (be the winner in a contest or competition; be victo- rious) 2. acquire, win, gain (win something through one’s efforts) 3. gain, advance, win, pull ahead, make headway, get ahead, gain ground (obtain advantages, such as points, etc.) 4. succeed, win, come through, bring home the bacon, de- liver the goods (attain success or reach a desired goal) In this context, many annotators would agree that the athlete has both won an object (the gold metal itself) and won a competition (signified by the gold medal). Although contexts can be constructed to elicit only one of these senses, in the example above, a graded annotation best matches human perception. Graded word sense (GWS) annotation offers sig- nificant advantages for sense annotation with a fine- grained sense inventory. However, creating a suf- ficiently large annotated corpus for training super- vised GWS disambiguation models presents a sig- nificant …

Content truncated to 500 words