Robust Entity Clustering via Phylogenetic Inference

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 775–785, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Robust Entity Clustering via Phylogenetic Inference Nicholas Andrews and Jason Eisner and Mark Dredze Department of Computer Science and Human Language Technology Center of Excellence Johns Hopkins University 3400 N. Charles St., Baltimore, MD 21218 USA {noa,eisner,mdredze}@jhu.edu Abstract Entity clustering must determine when two named-entity mentions refer to the same entity. Typical approaches use a pipeline ar- chitecture that clusters the mentions using fixed or learned measures of name and con- text similarity. In this paper, we propose a model for cross-document coreference res- olution that achieves robustness by learn- ing similarity from unlabeled data. The generative process assumes that each entity mention arises from copying and option- ally mutating an earlier name from a sim- ilar context. Clustering the mentions into entities depends on recovering this copying tree jointly with estimating models of the mutation process and parent selection pro- cess. We present a block Gibbs sampler for posterior inference and an empirical evalu- ation on several datasets. 1 Introduction Variation poses a serious challenge for determin- ing who or what a name refers to. For instance, Wikipedia contains more than 100 variations of the name Barack Obama as redirects to the U.S. Presi- dent article, including: President Obama Barack H. Obama, Jr. Barak Obamba Barry Soetoro To relate different names, one solution is to use specifically tailored measures of name similarity such as Jaro-Winkler similarity (Winkler, 1999; Co- hen et al., 2003). This approach is brittle, however, and fails to adapt to the test data. Another option is to train a model like stochastic edit distance from known pairs of similar names (Ristad and Yian- ilos, 1998; Green et al., 2012), but this requires supervised data in the test domain. Even the best model of name similarity is not enough by itself, since two names that are similar— even identical—do not necessarily corefer. Docu- ment context is needed to determine whether they may be talking about two different people. In this paper, we propose a method for jointly (1) learning similarity between names and (2) clus- tering name mentions into entities, the two major components of cross-document coreference reso- lution systems (Baron and Freedman, 2008; Finin et al., 2009; Rao et al., 2010; Singh et al., 2011; Lee et al., 2012; Green et al., 2012). Our model is an evolutionary generative process based on the name variation model of Andrews et al. (2012), which stipulates that names are often copied from previously generated names, perhaps with mutation (spelling edits). This can deduce that rather than being names for different entities, Barak Obamba and Barock obama more likely arose from the fre- quent name Barack Obama as a common ancestor, which accounts for most of their letters. This can also relate seemingly dissimilar names via multiple steps in the generative process: Taylor Swift → T-Swift → T-Swizzle Our model learns without supervision that these …

Content truncated to 500 words