Syntactic Parsing for Ranking-Based Coreference Resolution

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 465–473, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Syntactic Parsing for Ranking-Based Coreference Resolution Altaf Rahman and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {altaf,vince}@hlt.utdallas.edu Abstract Recent research efforts have led to the de- velopment of a state-of-the-art supervised coreference model, the cluster-ranking model. However, it is not clear whether the features that have been shown to be useful when employed in traditional coreference models will fare similarly when used in combination with this new model. Rather than merely re-evaluate them using the cluster-ranking model, we examine two in- teresting types of features derived from syntactic parses, tree-based features and path-based features, and discuss the chal- lenges involved in employing them in the cluster-ranking model. Results on a set of Switchboard dialogues show their effec- tiveness in improving the cluster-ranking model: using them to augment a baseline coreference feature set yields a 8.6–11.7% reduction in relative error. 1 Introduction Coreference resolution is the task of determining which noun phrases (NPs) in a text or dialogue refer to the same real-world entity. According to Webber (1979), coreference resolution can be de- composed into two complementary subtasks: “(1) identifying what a text potentially makes avail- able for anaphoric reference and (2) constraining the candidate set of a given anaphoric expression down to one possible choice”. These two subtasks are commonly known as anaphoricity determina- tion and anaphora resolution, both of which have recently been tackled using machine learning tech- niques. More specifically, anaphoricity determina- tion is typically tackled by training an anaphoric- ity classifier, which determines whether an NP is anaphoric or not (e.g., Poesio et al. (2004), Zhou and Kong (2009)). If so, the NP is passed to the second component, the resolution system, which identifies an antecedent for the NP. This resolver is typically implemented by training a mention-pair (MP) model, which is a binary classifier that deter- mines whether a pair of NPs are co-referring or not (e.g., Soon et al. (2001), Ng and Cardie (2002b)). While this architecture is popularly adopted by coreference researchers and was implemented even within recently developed coreference re- solvers (e.g., Bengtson and Roth (2008), Stoy- anov et al. (2009)), neither the architecture it- self nor its aforementioned implementation is sat- isfactory for at least two reasons. First, in this pipeline architecture, anaphoricity determination is performed prior to coreference resolution, so er- rors in anaphoricity determination can propagate to the downstream coreference component and ad- versely affect its performance (Ng and Cardie, 2002a). Second, the MP coreference model is fun- damentally weak in that (1) the information ex- tracted from two NPs may not be sufficient for making an informed coreference decision and (2) since the model is trained to compare the NP to be resolved (henceforth the active NP) against a can- didate antecedent, it only determines how good the candidate is relative to the active NP, not how good …

Content truncated to 500 words