Prune-and-Score: Learning for Greedy Coreference Resolution

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2115–2126, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Prune-and-Score: Learning for Greedy Coreference Resolution Chao Ma, Janardhan Rao Doppa†, J. Walker Orr, Prashanth Mannem Xiaoli Fern, Tom Dietterich and Prasad Tadepalli School of Electrical Engineering and Computer Science, Oregon State University {machao,orr,mannemp,xfern,tgd,tadepall}@eecs.oregonstate.edu † School of Electrical Engineering and Computer Science, Washington State University jana@eecs.wsu.edu 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 corefer- ence decision: a pruning function that prunes bad coreference decisions from fur- ther consideration, and a scoring function that then selects the best among the re- maining decisions. Our framework re- duces 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 sever- al state-of-the-art approaches on multiple benchmark corpora including OntoNotes. 1 Introduction Coreference resolution is the task of clustering a set of mentions in the text such that all mentions in the same cluster refer to the same entity. It is one of the first stages in deep language understanding and has a big potential impact on the rest of the stages. Several of the state-of-the-art approaches learn a scoring function defined over mention pair, cluster-mention or cluster-cluster pair to guide the coreference decision-making process (Daum´e II- I, 2006; Bengtson and Roth, 2008; Rahman and Ng, 2011b; Stoyanov and Eisner, 2012; Chang et al., 2013; Durrett et al., 2013; Durrett and Klein, 2013). One common and persistent problem with these approaches is that the scoring function has to make all the coreference decisions, which leads to a highly non-realizable learning problem. Inspired by the recent success of the HC-Search Framework (Doppa et al., 2014a) for studying a variety of structured prediction problems (Lam et al., 2013; Doppa et al., 2014c), we study a novel approach for search-based coreference resolution called Prune-and-Score. HC-Search is a divide- and-conquer solution that learns multiple compo- nents with pre-defined roles, and each of them contribute towards the overall goal by making the role of the other components easier. The HC- Search framework operates in the space of com- plete outputs, and relies on the loss function which is only defined on the complete outputs to drive it- s learning. Unfortunately, this method does not work for incremental coreference resolution since the search space for coreference resolution con- sists of partial outputs, i.e., a set of mentions only some of which have been clustered so far. We develop an alternative framework to HC- Search that allows us to effectively learn from par- tial output spaces and apply it to greedy corefer- ence resolution. The key idea of our work is to address the problem of non-realizability of the s- coring function by …

Content truncated to 500 words