Phrase Structures and Dependencies for End-to-End Coreference Resolution

We present experiments in data-driven coreference resolution comparing the effect of different syntactic representations provided as features in the coreference classification step: no syntax, phrase structure representations, dependency representations, and combinations of the representation types. We compare the end-to-end performance of a parametrized state-of-the-art coreference resolution system on the English data from the CoNLL 2012 shared task. On their own, phrase structures are more useful than dependencies, but the combinations yield highest performance and a significant improvement on the resolution of pronouns. Enriching phrase structure with dependency trees obtained from an independent parser is most helpful, but an extension of the predicted phrase structure using just pattern-based phraseto-dependency conversion seems to provide signals for the machine learning that cannot be distilled from phrase structure alone (despite intense feature selection). This is an interesting result for a highly configurational language: It is easier to learn generalizations over grammatical constraints on coreference when grammatical relations are explicitly provided.