A Lazy Learning Model for Entity Linking using Query-Specific Information

Proceedings of COLING 2012: Technical Papers, pages 3089–3104, COLING 2012, Mumbai, December 2012. A Lazy Learning Model for Entity Linking Using Query-Specific Information WeiZhang1 JianSu2 ChewLimTan1 YunboCao3 ChinYewLin3 (1) National University of Singapore (2) Institute for Infocomm Research, Singapore (3) Microsoft Research Asia {z-wei,tancl}@comp.nus.edu.sg, sujian@i2r.a-star.edu.sg, {Yunbo.Cao,cyl}@microsoft.com Abstract Entity linking disambiguates a mention of an entity in text to a Knowledge Base (KB). Most previous studies disambiguate a mention of a name (e.g.“AZ”) based on the distribution knowledge learned from labeled instances, which are related to other names (e.g.“Hoffman”,“Chad Johnson”, etc.). The gaps among the distributions of the instances related to different names hinder the further improvement of the previous approaches. This paper proposes a lazy learning model, which allows us to improve the learning process with the distribution information specific to the queried name (e.g.“AZ”). To obtain this distribution information, we automatically label some relevant instances for the queried name leveraging its unambiguous synonyms. Besides, another advantage is that our approach still can benefit from the labeled data related to other names (e.g.“Hoffman”,“Chad Johnson”, etc.), because our model is trained on both the labeled data sets of queried and other names by mining their shared predictive structure. Keywords: Entity Linking, Lazy Learning, Query-Specific Information. 3089 1 Introduction Recently, more and more knowledge bases (KB) which contain rich knowledge about the world’s entities such as Wikipedia 1, OpenCyc 2 and KIM 3 (Popov et al., 2004) have become available. These knowledge bases have been shown to form a valuable component for many natural language processing tasks such as knowledge base population (Ji and Grishman, 2011), text classification (Wang and Domeniconi, 2008), and cross-document coreference (Finin et al., 2009). However, to be able to utilize or enrich these KB resources, the applications usually require linking the mentions of entities in text to their corresponding entries in the knowledge bases, which is called entity linking task and has been proposed and studied in Text Analysis Conference (TAC) since 2009 (McNamee and Dang, 2009). Given a mention of an entity in text and a KB, entity linking is to link the mention to its corresponding entry in KB. The major challenges of this task are name variation and name ambiguity. Name variation refers to the case that more than one name variation such as alias, misspelling and acronym refers to the same entity. For example, both “48th State” and “The Grand Canyon State” refer to state of Arizona, U.S.. Name ambiguity refers to the case that more than one entity shares the same name. For example, “AZ” may refer to state of Arizona, the Italian airline Alitalia, the country Azerbaijan, or other entries in KB that have the same name. Most previous studies on entity linking used annotated data to learn a classifier or ranker (Dredze et al., 2010; Lehmann et al., 2010; Zheng et al., 2010; Zhang et al., 2010; Ploch, 2011; Ratinov et al., 2011) or to estimate parameters (Gottipati and Jiang, 2011; Han and Sun, 2011). Besides, from the analysis by …

Content truncated to 500 words