Coreference Resolution in a Modular, Entity-Centered Model

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 385–393, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Coreference Resolution in a Modular, Entity-Centered Model Aria Haghighi Computer Science Division University of California, Berkeley aria42@cs.berkeley.edu Dan Klein Computer Science Division University of California, Berkeley klein@cs.berkeley.edu Abstract Coreference resolution is governed by syntac- tic, semantic, and discourse constraints. We present a generative, model-based approach in which each of these factors is modularly en- capsulated and learned in a primarily unsu- pervised manner. Our semantic representation first hypothesizes an underlying set of latent entity types, which generate specific entities that in turn render individual mentions. By sharing lexical statistics at the level of abstract entity types, our model is able to substantially reduce semantic compatibility errors, result- ing in the best results to date on the complete end-to-end coreference task. 1 Introduction Coreference systems exploit a variety of informa- tion sources, ranging from syntactic and discourse constraints, which are highly configurational, to se- mantic constraints, which are highly contingent on lexical meaning and world knowledge. Perhaps be- cause configurational features are inherently easier to learn from small data sets, past work has often emphasized them over semantic knowledge. Of course, all state-of-the-art coreference systems have needed to capture semantic compatibility to some degree. As an example of nominal headword compatibility, a “president” can be a “leader” but cannot be not an “increase.” Past systems have of- ten computed the compatibility of specific headword pairs, extracted either from lexical resources (Ng, 2007; Bengston and Roth, 2008; Rahman and Ng, 2009), web statistics (Yang et al., 2005), or sur- face syntactic patterns (Haghighi and Klein, 2009). While the pairwise approach has high precision, it is neither realistic nor scalable to explicitly enumerate all pairs of compatible word pairs. A more compact approach has been to rely on named-entity recog- nition (NER) systems to give coarse-grained entity types for each mention (Soon et al., 1999; Ng and Cardie, 2002). Unfortunately, current systems use small inventories of types and so provide little con- straint. In general, coreference errors in state-of-the- art systems are frequently due to poor models of se- mantic compatibility (Haghighi and Klein, 2009). In this work, we take a primarily unsupervised ap- proach to coreference resolution, broadly similar to Haghighi and Klein (2007), which addresses this is- sue. Our generative model exploits a large inven- tory of distributional entity types, including standard NER types like PERSON and ORG, as well as more refined types like WEAPON and VEHICLE. For each type, distributions over typical heads, modifiers, and governors are learned from large amounts of unla- beled data, capturing type-level semantic informa- tion (e.g. “spokesman” is a likely head for a PER- SON). Each entity inherits from a type but captures entity-level semantic information (e.g. “giant” may be a likely head for the Microsoft entity but not all ORGs). Separately from the type-entity semantic module, a log-linear discourse model captures con- figurational effects. Finally, a …

Content truncated to 500 words