A joint inference architecture for global coreference clustering with anaphoricity

Abstract

We present an architecture for coreference resolution based on joint inference over anaphoricity and coreference, using Markov Logic Networks. Mentions are discriminatively clustered with discourse entities established by an anaphoricity classifier. Our entity-based coreference architecture is realized in a joint inference setting to compensate for erroneous anaphoricity classifications and avoids local coreference misclassifications through global consistency constraints. Defining pairwise coreference features in a global setting achieves an efficient entity-based perspective. With a small feature set we obtain a performance of 63.56% (gold mentions) on the official CoNLL 2012 data set. © 2013 Springer-Verlag.