Sense Clustering Using Wikipedia

Proceedings of Recent Advances in Natural Language Processing, pages 164–171, Hissar, Bulgaria, 7-13 September 2013. Sense Clustering Using Wikipedia Bharath Dandala, Chris Hokamp and Rada Mihalcea University of North Texas bharathdandala@gmail.com chris.hokamp@gmail.com rada@cs.unt.edu Razvan C. Bunescu Ohio University bunescu@ohio.edu Abstract In this paper, we propose a novel method for generating a coarse-grained sense in- ventory from Wikipedia using a ma- chine learning framework. Structural and content-based features are employed to in- duce clusters of articles representative of a word sense. Additionally, multilingual features are shown to improve the clus- tering accuracy, especially for languages that are less comprehensive than English. We show the effectiveness of our clus- tering methodology by testing it against both manually and automatically anno- tated datasets. 1 Introduction The granularity of word sense repositories has been recognized as an important factor in the de- velopment of annotated datasets for Word Sense Disambiguation (WSD) (Snow et al., 2007), with significant impacts upon both the performance of automatic WSD systems and their utility for downstream applications. Previous work on man- ual sense annotations with respect to WordNet has revealed low levels of agreement between human annotators, ranging between 65% (Chklovski and Mihalcea, 2002) and 72% (Snyder and Palmer, 2004), which is a clear indicator of very fine- grained word senses that are difficult to differen- tiate, even for humans. To achieve the sense granularity appropriate for WSD, word senses that are closely related in meaning are grouped together in a sense cluster- ing step. While this task was originally defined in relation to more traditional sense inventories, such as WordNet (Hovy et al., 2006; Mihalcea and Moldovan, 2001) or the Oxford dictionary (Nav- igli, 2006), newer user-contributed sense inven- tories such as Wikipedia or Wiktionary are also quickly expanding and refining the senses defined for a word, thus pointing to the need of sense clus- tering for coarser word sense distinctions. In this paper, we specifically focus on the task of sense clustering over Wikipedia senses. Wikipedia has been recently recognized as a rich resource for WSD (Bunescu and Pasca, 2006; Mihalcea, 2007; Mihalcea and Csomai, 2007; Milne and Witten, 2008), offering a significantly increased coverage of word meanings relative to established reposito- ries such as WordNet or Roget. At the same time, WSD systems using Wikipedia have been shown to obtain comparable or even increased disam- biguation precision. While earlier work on WSD using the 2007 version of Wikipedia reported an average of three senses per word for a dataset of 30 nouns (Mihalcea, 2007), more recent work on the same dataset using the 2012 version of Wikipedia has shown a significant increase to an average of nine senses per word (Dandala et al., 2012). For instance, the noun “paper”, which used to have five different senses, now has ten senses; similarly, the noun “bar”, which previously had ten senses, now has 23 senses. The accuracy of a WSD sys- tem on the same set of 30 nouns dropped from an average of 85% when using Wikipedia …

Content truncated to 500 words