Prune-and-Score: Learning for Greedy Coreference Resolution

Abstract

We propose a novel search-based approach for greedy coreference resolution, where the mentions are processed in order and added to previous coreference clusters. Our method is distinguished by the use of two functions to make each coreference decision: a pruning function that prunes bad coreference decisions from further consideration, and a scoring function that then selects the best among the remaining decisions. Our framework reduces learning of these functions to rank learning, which helps leverage powerful off-the-shelf rank-learners. We show that our Prune-and-Score approach is superior to using a single scoring function to make both decisions and outperforms several state-of-the-art approaches on multiple benchmark corpora including OntoNotes.

Entity-Centric Coreference Resolution with Model Stacking

“…The majority of useful features for coreference systems operate on pairs of mentions (in one of our experiments we show the addition of classic entity-level features does not improve our system), but incremental coreference systems must make decisions involving many mention pairs. Other incremental coreference systems either incorporate features from a single pair (Stoyanov and Eisner, 2012) or average features across all pairs in the involved clusters (Ma et al, 2014). Our system instead combines information from the involved mention pairs in a variety of ways with with higher order features produced from the scores of mention pair models.…”

“…In Table 3 we compare the results of our system with the following state-of-the-art approaches: the JOINT and INDEP models of the Berkeley system (Durrett and Klein, 2014) (the JOINT model jointly does NER and entity linking along with coreference); the Prune-and-Score system (Ma et al, 2014); the HOTCoref system (Björkelund and Kuhn, 2014); the CPL 3 M sytem (Chang et al, 2013); and Fernandes et al We use the full entitycentric clustering algorithm drawing upon scores from both pairwise models. We do not make use of agreement features, as these did not increase accuracy and complicate the system.…”

“…Imitation learning has been employed to train coreference resolvers on trajectories of decisions similar to those that would be seen at test-time by Daumé et al (2005) and Ma et al (2014). Other works use structured perceptron models for the same purpose (Stoyanov and Eisner, 2012;Fernandes et al, 2012;Björkelund and Kuhn, 2014).…”

“…For example, finding that Clinton and she corefer makes it more likely that Clinton corefers with Hillary Clinton than Bill Clinton due to gender agreement constraints. Such information has been incorporated successfully into entity-centric coreference systems that build up coreference clusters incrementally, using the information from the partially completed coreference chains produced so far to guide later decisions (Raghunathan et al, 2010;Stoyanov and Eisner, 2012;Ma et al, 2014).…”

Search Space Pruning: A Simple Solution for Better Coreference Resolvers

“…Among those, anaphoricity detection is the most popular method (e.g. Ng and Cardie (2002), Denis and Baldridge (2007), Ng (2009), Zhou andKong (2009), Durrett andKlein (2013), Martschat and Strube (2015), Wiseman et al (2015), Peng et al (2015), and Lassalle and Denis (2015)), while singleton detection is a more recent method (Recasens et al, 2013;Ma et al, 2014;Marneffe et al, 2015).…”