Specialized Models and Ranking for Coreference Resolution

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 660–669, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Specialized models and ranking for coreference resolution Pascal Denis ALPAGE Project Team INRIA Rocquencourt F-78153 Le Chesnay, France pascal.denis@inria.fr Jason Baldridge Department of Linguistics University of Texas at Austin Austin, TX 78712-0198, USA jbaldrid@mail.utexas.edu Abstract This paper investigates two strategies for im- proving coreference resolution: (1) training separate models that specialize in particu- lar types of mentions (e.g., pronouns versus proper nouns) and (2) using a ranking loss function rather than a classification function. In addition to being conceptually simple, these modifications of the standard single-model, classification-based approach also deliver sig- nificant performance improvements. Specifi- cally, we show that on the ACE corpus both strategies produce f-score gains of more than 3% across the three coreference evaluation metrics (MUC, B3, and CEAF). 1 Introduction Coreference resolution is the task of partitioning a set of entity mentions in a text, where each par- tition corresponds to some entity in an underlying discourse model. While early machine learning ap- proaches for the task relied on local, discriminative classifiers (Soon et al., 2001; Ng and Cardie, 2002b; Morton, 2000; Kehler et al., 2004), more recent ap- proaches use joint and/or global models (McCallum and Wellner, 2004; Ng, 2004; Daum´e III and Marcu, 2005; Denis and Baldridge, 2007a). This shift im- proves performance, but the systems are consider- ably more complex and often less efficient. Here, we explore two simple modifications of the first type of approach that yield performance gains which are comparable, and sometimes better, to those obtained with these more complex systems. These modifica- tions involve: (i) the use of rankers instead of clas- sifiers, and (ii) the use of linguistically motivated, specialized models for different types of mentions. Ranking models provide a theoretically more ad- equate and empirically better alternative approach to pronoun resolution than standard classification- based approaches (Denis and Baldridge, 2007b). In essence, ranking models directly capture during training the competition among potential antecedent candidates, instead of considering them indepen- dently. This gives the ranker additional discrimina- tive power and in turn better antecedent selection ac- curacy. Here, we show that ranking is also effective for the wider task of coreference resolution. Coreference resolution involves several different types of anaphoric expressions: third-person pro- nouns, speech pronouns (i.e., first and second person pronouns), proper names, definite descriptions and other types of nominals (e.g., anaphoric uses of in- definite, quantified, and bare noun phrases). Differ- ent anaphoric expressions exhibit different patterns of resolution and are sensitive to different factors (Ariel, 1988; van der Sandt, 1992; Gundel et al., 1993), yet most machine learning approaches have ignored these differences and handle these different phenomena with a single, monolithic model. A few exceptions are worth noting. Morton (2000) and Ng (2005b) propose different classifiers models for dif- ferent NPs for coreference resolution and pronoun resolution, respectively. Other partially capture the differential preferences between different anaphors via different sample selection strategies …

Content truncated to 500 words