Local and Global Algorithms for Disambiguation to Wikipedia

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 1375–1384, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Local and Global Algorithms for Disambiguation to Wikipedia Lev Ratinov 1 Dan Roth1 Doug Downey2 Mike Anderson3 1University of Illinois at Urbana-Champaign {ratinov2|danr}@uiuc.edu 2Northwestern University ddowney@eecs.northwestern.edu 3Rexonomy mrander@gmail.com Abstract Disambiguating concepts and entities in a con- text sensitive way is a fundamental problem in natural language processing. The compre- hensiveness of Wikipedia has made the on- line encyclopedia an increasingly popular tar- get for disambiguation. Disambiguation to Wikipedia is similar to a traditional Word Sense Disambiguation task, but distinct in that the Wikipedia link structure provides addi- tional information about which disambigua- tions are compatible. In this work we analyze approaches that utilize this information to ar- rive at coherent sets of disambiguations for a given document (which we call “global” ap- proaches), and compare them to more tradi- tional (local) approaches. We show that previ- ous approaches for global disambiguation can be improved, but even then the local disam- biguation provides a baseline which is very hard to beat. 1 Introduction Wikification is the task of identifying and link- ing expressions in text to their referent Wikipedia pages. Recently, Wikification has been shown to form a valuable component for numerous natural language processing tasks including text classifica- tion (Gabrilovich and Markovitch, 2007b; Chang et al., 2008), measuring semantic similarity between texts (Gabrilovich and Markovitch, 2007a), cross- document co-reference resolution (Finin et al., 2009; Mayfield et al., 2009), and other tasks (Kulkarni et al., 2009). Previous studies on Wikification differ with re- spect to the corpora they address and the subset of expressions they attempt to link. For exam- ple, some studies focus on linking only named en- tities, whereas others attempt to link all “interest- ing” expressions, mimicking the link structure found in Wikipedia. Regardless, all Wikification systems are faced with a key Disambiguation to Wikipedia (D2W) task. In the D2W task, we’re given a text along with explicitly identified substrings (called mentions) to disambiguate, and the goal is to out- put the corresponding Wikipedia page, if any, for each mention. For example, given the input sen- tence “I am visiting friends in <Chicago>,” we output http://en.wikipedia.org/wiki/Chicago – the Wikipedia page for the city of Chicago, Illinois, and not (for example) the page for the 2002 film of the same name. Local D2W approaches disambiguate each men- tion in a document separately, utilizing clues such as the textual similarity between the document and each candidate disambiguation’s Wikipedia page. Recent work on D2W has tended to focus on more sophisticated global approaches to the problem, in which all mentions in a document are disambiguated simultaneously to arrive at a coherent set of dis- ambiguations (Cucerzan, 2007; Milne and Witten, 2008b; Han and Zhao, 2009). For example, if a mention of “Michael Jordan” refers to the computer scientist rather than the basketball player, then we would expect a mention of “Monte Carlo” in the same document to …

Content truncated to 500 words