Collaborative Partitioning for Coreference Resolution

Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 47–57, Vancouver, Canada, August 3 - August 4, 2017. c⃝2017 Association for Computational Linguistics Collaborative Partitioning for Coreference Resolution Olga Uryupina♦ and Alessandro Moschitti ♦DISI, University of Trento 38123 Povo (TN), Italy Qatar Computing Research Institute, HBKU, 34110, Doha, Qatar {uryupina,amoschitti}@gmail.com Abstract This paper presents a collaborative par- titioning algorithm—a novel ensemble- based approach to coreference resolution. Starting from the all-singleton partition, we search for a solution close to the en- semble’s outputs in terms of a task-specific similarity measure. Our approach assumes a loose integration of individual compo- nents of the ensemble and can therefore combine arbitrary coreference resolvers, regardless of their models. Our experi- ments on the CoNLL dataset show that collaborative partitioning yields results su- perior to those attained by the individual components, for ensembles of both strong and weak systems. Moreover, by applying the collaborative partitioning algorithm on top of three state-of-the-art resolvers, we obtain the second-best coreference per- formance reported so far in the literature (MELA v08 score of 64.47). 1 Introduction Coreference resolution has been one of the key areas of NLP for several decades. Major mod- eling breakthroughs have been achieved, not sur- prisingly, following three successful shared tasks, such as MUC (Hirschman and Chinchor, 1997), ACE (Doddington et al., 2004) and, most recently, CoNLL (Pradhan et al., 2011; Pradhan et al., 2012). As of today, several high-performing sys- tems are available publicly and, in addition, novel algorithms are being proposed regularly, even if without any code release. Our study aims at mak- ing a good use of these resources through a novel ensemble resolution method. Coreference is a heterogeneous task that re- quires a combination of accurate and robust pro- cessing for relatively easy cases (e.g., name- matching) with very complex modeling of diffi- cult cases (e.g., nominal anaphora or some types of pronouns). The general feeling in the commu- nity is that we are currently approaching the upper bound for the easy cases and our next step should involve more complex resolution. If true, this means that most state-of-the-art systems should produce very similar outputs: correctly resolving easy anaphora and failing on less trivial examples. Table 1 scores the outputs of the three best sys- tems from the CoNLL-2012 shared task against each other. As it can be seen, the three systems are rather different, each of them being only slightly closer to each other than to the gold key.1 This suggests that a meta-algorithm could merge their outputs in an intelligent way, combining the cor- rect decisions of individual systems to arrive at a superior partition. Although several coreference resolution toolk- its exist for over a decade, to our knowledge, there have been no attempts at trying to merge their out- puts. The very few ensemble methods reported in the literature focus on combining several resolu- tion strategies within the same system. Following the success of the CoNLL shared task (Pradhan et al., 2011; Pradhan et al., …

Content truncated to 500 words