The More Antecedents, the Merrier: Resolving Multi-Antecedent Anaphors

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 2287–2296, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics The More Antecedents, the Merrier: Resolving Multi-Antecedent Anaphors Hardik Vala1, Andrew Piper2, Derek Ruths1 1School of Computer Science, 2Dept. of Languages, Literartures, & Cultures McGill University Montreal, Canada hardik.vala@mail.mcgill.ca, {andrew.piper, derek.ruths}@mcgill.ca Abstract Anaphor resolution is an important task in NLP with many applications. De- spite much research effort, it remains an open problem. The difficulty of the prob- lem varies substantially across different sub-problems. One sub-problem, in par- ticular, has been largely untouched by prior work despite occurring frequently throughout corpora: the anaphor that has multiple antecedents, which here we call multi-antecedent anaphors or m- anaphors. Current coreference resolvers restrict anaphors to at most a single an- tecedent. As we show in this paper, re- laxing this constraint poses serious prob- lems in coreference chain-building, where each chain is intended to refer to a single entity. This work provides a formaliza- tion of the new task with preliminary in- sights into multi-antecedent noun-phrase anaphors, and offers a method for resolv- ing such cases that outperforms a number of baseline methods by a significant mar- gin. Our system uses local agglomerative clustering on candidate antecedents and an existing coreference system to score clus- ters to determine which cluster of men- tions is antecedent for a given anaphor. When we augment an existing coreference system with our proposed method, we ob- serve a substantial increase in performance (0.6 absolute CoNLL F1) on an annotated corpus. 1 Introduction Anaphor resolution is a very difficult task in Nat- ural Language Understanding, involving the com- plex interaction of discourse cues, syntactic rules, and semantic phenomena. It is closely related to the task of coreference resolution (Van Deemter and Kibble, 2000), for which a myriad of solu- tions have been proposed (Clark and Manning, 2015; Peng et al., 2015; Wiseman et al., 2015; Bj¨orkelund and Farkas, 2012; Lee et al., 2011; Stoyanov et al., 2010; Ng, 2008; Bergsma and Lin, 2006; Soon et al., 2001). However, given the complexity of the problem, a comprehensive approach remains elusive. The difficulty varies drastically across different cases (proper nouns, pronouns, gerunds, etc.), each of which involves different assumptions about and models of vari- ous linguistic phenomena (e.g., vocabulary, syn- tax, and semantics). As a result, state-of-the- art systems yield varying performance across sub- problems (Mitkov, 2014; Kummerfeld and Klein, 2013; Bj¨orkelund and Nugues, 2011; Recasens and Hovy, 2009; Stoyanov et al., 2009; Bengtson and Roth, 2008; Van Deemter and Kibble, 2000; Ng and Cardie, 2002b; Kameyama, 1997). To avoid the complexity of the overarching res- olution task, many current systems — whether learning-based (Clark and Manning, 2015; Peng et al., 2015; Wiseman et al., 2015; Durrett and Klein, 2013; Bj¨orkelund and Farkas, 2012) or rule-based (Lee et al., 2011) — focus on a restricted ver- sion of the problem, where candidate anaphors are linked to at most one antecedent, from which coreference chains are built …

Content truncated to 500 words