Coarse to Fine Grained Sense Disambiguation in Wikipedia

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 22–31, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics Coarse to Fine Grained Sense Disambiguation in Wikipedia Hui Shen School of EECS Ohio University Athens, OH 45701, USA hui.shen.1@ohio.edu Razvan Bunescu School of EECS Ohio University Athens, OH 45701, USA bunescu@ohio.edu Rada Mihalcea Department of CSE University of North Texas Denton, TX 76203, USA rada@cs.unt.edu Abstract Wikipedia articles are annotated by volunteer contributors with numerous links that connect words and phrases to relevant titles. Links to general senses of a word are used concur- rently with links to more specific senses, with- out being distinguished explicitly. We present an approach to training coarse to fine grained sense disambiguation systems in the presence of such annotation inconsistencies. Experi- mental results show that accounting for anno- tation ambiguity in Wikipedia links leads to significant improvements in disambiguation. 1 Introduction and Motivation The vast amount of world knowledge available in Wikipedia has been shown to benefit many types of text processing tasks, such as coreference res- olution (Ponzetto and Strube, 2006; Haghighi and Klein, 2009; Bryl et al., 2010; Rahman and Ng, 2011), information retrieval (Milne, 2007; Li et al., 2007; Potthast et al., 2008; Cimiano et al., 2009), or question answering (Ahn et al., 2004; Kaisser, 2008; Ferrucci et al., 2010). In particular, the user contributed link structure of Wikipedia has been shown to provide useful supervision for training named entity disambiguation (Bunescu and Pasca, 2006; Cucerzan, 2007) and word sense disambigua- tion (Mihalcea, 2007; Ponzetto and Navigli, 2010) systems. Articles in Wikipedia often contain men- tions of concepts or entities that already have a cor- responding article. When contributing authors men- tion an existing Wikipedia entity inside an article, they are required to link at least its first mention to the corresponding article, by using links or piped links. Consider, for example, the following Wiki source annotations: The [[capital city|capital]] of Georgia is [[Atlanta]]. The bracketed strings iden- tify the title of the Wikipedia articles that describe the corresponding named entities. If the editor wants a different string displayed in the rendered text, then the alternative string is included in a piped link, af- ter the title string. Based on these Wiki processing rules, the text that is rendered for the aforementioned example is: The capital of Georgia is Atlanta. Since many words and names mentioned in Wikipedia articles are inherently ambiguous, their corresponding links can be seen as a useful source of supervision for training named entity and word sense disambiguation systems. For example, Wikipedia contains articles that describe possible senses of the word “capital”, such as CAPITAL CITY, CAPITAL (ECONOMICS), FINANCIAL CAPITAL, or HUMAN CAPITAL, to name only a few. When dis- ambiguating a word or a phrase in Wikipedia, a con- tributor uses the context to determine the appropriate Wikipedia title to include in the link. In the exam- ple above, the editor of the …

Content truncated to 500 words