Global Inference for Bridging Anaphora Resolution

Proceedings of NAACL-HLT 2013, pages 907–917, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Global Inference for Bridging Anaphora Resolution Yufang Hou1, Katja Markert2, Michael Strube1 1 Heidelberg Institute for Theoretical Studies gGmbH, Heidelberg, Germany (yufang.hou|michael.strube)@h-its.org 2School of Computing, University of Leeds, UK scskm@leeds.ac.uk Abstract We present the first work on antecedent se- lection for bridging resolution without restric- tions on anaphor or relation types. Our model integrates global constraints on top of a rich local feature set in the framework of Markov logic networks. The global model improves over the local one and both strongly outper- form a reimplementation of prior work. 1 Introduction Identity coreference is a relatively well understood and well-studied instance of entity coherence. How- ever, entity coherence can rely on more complex, lexico-semantic, frame or encyclopedic relations than identity. Anaphora linking distinct entities or events this way are called bridging or associative anaphora and have been widely discussed in the lin- guistic literature (Clark, 1975; Prince, 1981; Gundel et al., 1993).1 In Example 1, the phrases the win- dows, the carpets and walls can be felicitously used because they are semantically related via a part-of relation to their antecedent the Polish center.2 (1) . . . as much as possible of the Polish center will be made from aluminum, steel and glass recycled from Warsaw’s abundant rubble. . . . The windows will open. The carpets won’t be glued down and walls will be coated with non-toxic finishes. 1Poesio and Vieira (1998) include cases where antecedent and anaphor are coreferent but do not share the same head noun. We restrict bridging to non-coreferential cases. We also exclude comparative anaphora (Modjeska et al., 2003) 2Examples are from OntoNotes (Weischedel et al., 2011). Bridging anaphora are typed in boldface; antecedents in italics. Bridging is frequent amounting to between 5% (Gardent and Manu´elian, 2005) and 20% (Caselli and Prodanof, 2006) of definite descriptions (both studies limited to NPs starting with the or non- English equivalents). Bridging resolution is needed to fill gaps in entity grids based on coreference only (Barzilay and Lapata, 2008). Example 1 does not ex- hibit any coreferential entity coherence. Coherence can only be established when the bridging anaphora are resolved. Bridging resolution may also be im- portant for textual entailment (Mirkin et al., 2010). Bridging resolution can be divided into two tasks, recognizing that a bridging anaphor is present and finding the correct antecedent among a list of candi- dates. These two tasks have frequently been handled in a pipeline with most research concentrating on an- tecedent selection only. We also handle only the task of antecedent selection. Previous work on antecedent selection for bridg- ing anaphora is restricted. It makes strong untested assumptions about bridging anaphora types or rela- tions, limiting it to definite NPs (Poesio and Vieira, 1998; Poesio et al., 2004; Lassalle and Denis, 2011) or to part-of relations between anaphor and an- tecedent (Poesio et al., 2004; Markert et al., 2003; Lassalle and Denis, 2011). We break new …

Content truncated to 500 words