Relational Inference for Wikification

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1787–1796, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Relational Inference for Wikification Xiao Cheng Dan Roth Department of Computer Science University of Illinois at Urbana-Champaign Urbana, IL 61801 {cheng88,danr}@illinois.edu Abstract Wikification, commonly referred to as Disam- biguation to Wikipedia (D2W), is the task of identifying concepts and entities in text and disambiguating them into the most specific corresponding Wikipedia pages. Previous ap- proaches to D2W focused on the use of lo- cal and global statistics over the given text, Wikipedia articles and its link structures, to evaluate context compatibility among a list of probable candidates. However, these meth- ods fail (often, embarrassingly), when some level of text understanding is needed to sup- port Wikification. In this paper we introduce a novel approach to Wikification by incorpo- rating, along with statistical methods, richer relational analysis of the text. We provide an extensible, efficient and modular Integer Lin- ear Programming (ILP) formulation of Wik- ification that incorporates the entity-relation inference problem, and show that the ability to identify relations in text helps both candi- date generation and ranking Wikipedia titles considerably. Our results show significant im- provements in both Wikification and the TAC Entity Linking task. 1 Introduction Wikification (D2W), the task of identifying concepts and entities in text and disambiguating them into their corresponding Wikipedia page, is an important step toward supporting deeper textual understand- ing, by augmenting the ability to ground text in ex- isting knowledge and facilitating knowledge expan- sion. D2W has been studied extensively recently (Cucerzan, 2007; Mihalcea and Csomai, 2007; Milne and Witten, 2008; Ferragina and Scaiella, 2010; Ratinov et al., 2011) and has already found broad applications in NLP, Information Extraction, and Knowledge Acquisition from text, from coref- erence resolution (Ratinov and Roth, 2012) to entity linking and knowledge population (Ellis et al., 2011; Ji et al., 2010; Cucerzan, 2011). Given a document D containing a set of concept and entity mentions M ( referred to later as surface), the goal of Wikification is to find the most accurate mapping from mentions to Wikipedia titles T; this mapping needs to take into account our understand- ing of the text as well as background knowledge that is often needed to determine the most appropriate ti- tle. We also allow a special NIL title that captures all mentions that are outside Wikipedia. Earlier approaches treated this task as a word- sense disambiguation (WSD) problem, which was later enhanced with a certain level of global rea- soning, but essentially all approaches focused on generic statistical features in order to achieve robust disambiguation. It was shown that by disambiguat- ing to the most likely title for every surface, in- dependently maximizing the conditional probability Pr(title|surface), we already achieve a very com- petitive baseline on several Wikification datasets (Ratinov et al., 2011). This strong statistical baseline makes use of the relatively comprehensive coverage of the existing Wikipedia links from surface strings to Wikipedia titles. …

Content truncated to 500 words