Improving pairwise coreference models through feature space hierarchy learning

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 497–506, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Improving pairwise coreference models through feature space hierarchy learning Emmanuel Lassalle Alpage Project-team INRIA & Univ. Paris Diderot Sorbonne Paris Cit´e, F-75205 Paris emmanuel.lassalle@ens-lyon.org Pascal Denis Magnet Project INRIA Lille - Nord Europe Avenue Helo¨ıse, 59650 Villeneuve d’Ascq pascal.denis@inria.fr Abstract This paper proposes a new method for significantly improving the performance of pairwise coreference models. Given a set of indicators, our method learns how to best separate types of mention pairs into equivalence classes for which we con- struct distinct classification models. In ef- fect, our approach finds an optimal fea- ture space (derived from a base feature set and indicator set) for discriminating coref- erential mention pairs. Although our ap- proach explores a very large space of pos- sible feature spaces, it remains tractable by exploiting the structure of the hierar- chies built from the indicators. Our exper- iments on the CoNLL-2012 Shared Task English datasets (gold mentions) indicate that our method is robust relative to dif- ferent clustering strategies and evaluation metrics, showing large and consistent im- provements over a single pairwise model using the same base features. Our best system obtains a competitive 67.2 of aver- age F1 over MUC, B3, and CEAF which, despite its simplicity, places it above the mean score of other systems on these datasets. 1 Introduction Coreference resolution is the problem of partition- ing a sequence of noun phrases (or mentions), as they occur in a natural language text, into a set of referential entities. A common approach to this problem is to separate it into two modules: on the one hand, one defines a model for evaluating coreference links, in general a discriminative clas- sifier that detects coreferential mention pairs. On the other hand, one designs a method for group- ing the detected links into a coherent global out- put (i.e. a partition over the set of entity men- tions). This second step is typically achieved using greedy heuristics (McCarthy and Lehnert, 1995; Soon et al., 2001; Ng and Cardie, 2002; Bengston and Roth, 2008), although more so- phisticated clustering approaches have been used, too, such as cutting graph methods (Nicolae and Nicolae, 2006; Cai and Strube, 2010) and Integer Linear Programming (ILP) formulations (Klenner, 2007; Denis and Baldridge, 2009). Despite its simplicity, this two-step strategy remains competi- tive even when compared to more complex models utilizing a global loss (Bengston and Roth, 2008). In this kind of architecture, the performance of the entire coreference system strongly depends on the quality of the local pairwise classifier.1 Con- sequently, a lot of research effort on coreference resolution has focused on trying to boost the per- formance of the pairwise classifier. Numerous studies are concerned with feature extraction, typ- ically trying to enrich the classifier with more linguistic knowledge and/or more world knowl- edge (Ng and Cardie, 2002; Kehler et al., 2004; Ponzetto and Strube, 2006; Bengston and Roth, …

Content truncated to 500 words