Search Space Pruning: A Simple Solution for Better Coreference Resolvers

Proceedings of NAACL-HLT 2016, pages 1005–1011, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Search Space Pruning: A Simple Solution for Better Coreference Resolvers Nafise Sadat Moosavi and Michael Strube Heidelberg Institute for Theoretical Studies gGmbH Schloss-Wolfsbrunnenweg 35 69118 Heidelberg, Germany {nafise.moosavi|michael.strube}@h-its.org Abstract There is a significant gap between the per- formance of a coreference resolution system on gold mentions and on system mentions. This gap is due to the large and unbalanced search space in coreference resolution when using system mentions. In this paper we show that search space pruning is a simple but effi- cient way of improving coreference resolvers. By incorporating our pruning method in one of the state-of-the-art coreference resolution systems, we achieve the best reported over- all score on the CoNLL 2012 English test set. A version of our pruning method is available with the Cort coreference resolution source code. 1 Introduction Coreference resolution is the task of clustering refer- ring expressions in a text so that each resulting clus- ter represents an entity. It is a very challenging task in natural language processing and it is still far from being solved, i.e. the best reported overall CoNLL score on the CoNLL 2012 English test set is 63.39 (Wiseman et al., 2015). Text spans referring to an entity are called men- tions. Mentions are the primary objects in a corefer- ence resolution system. As with most previous work on coreference resolution, we only consider men- tions that are noun phrases. However, not all of the noun phrases are mentions. A noun phrase may not refer to any entity at all. The pronoun it in the sen- tence it is raining is an example of a non-referential noun phrase. Noun phrases which do refer to an en- tity (mentions) can be further divided into two cat- egories: mentions referring to entities which only appear once in the discourse (i.e. singletons), and mentions realizing entities that have been referred to more than once in the text (i.e. coreferent men- tions). Henceforth, we refer to both singletons and non-referential phrases as non-coreferent mentions. A large number of mentions that appear in a text are non-coreferent. For instance, more than 80% of mentions are singletons in the OntoNotes English development set (Marneffe et al., 2015). The latent ranking model is the best perform- ing model for coreference resolution to date (Wise- man et al., 2015; Martschat and Strube, 2015). If we use gold mentions, the latent ranking model of Martschat and Strube (2015) achieves an overall score of 80% on the CoNLL 2012 English test set. This result shows that once we have the ideal pruned search space, the ranking model with the current set of features is reasonably capable of finding corre- sponding entities of mentions. The substantial gap (17%) between the results of the gold mentions and system mentions implies that search space pruning is a promising direction for further improvements in coreference resolution. Marneffe et al. (2015) examine different search space pruning methods …

Content truncated to 500 words