Improve coreference resolution with parameter tunable anaphoricity identification and global optimization

Abstract

We build an anaphoric classifier with tunable parameters and realize a global connection between the classifier and coreference resolution. 60 features are used to build the anaphoric classifier. A corpus ratio control method is proposed and a "probability threshold" method is introduced to tune the precision and recall of the anaphoric classifier. The anaphoricity identification joints with the coreference resolution in a way of global optimization, and the parameters of anaphoricity identification are tuned according the result of coreference resolution. Maximum entropy is used for anaphoricity identification and coreference resolution with selected features. The results that combine the coreference resolution with the anaphoric classifier with different recall and precision are analyzed, and a comparison between our system and other coreference resolution systems is taken in the experiments analyze part. Our system improves the baseline coreference resolution system from 50.57 raise up to 53.35 on CoNLL'11 share tasks development data set. © 2012 Springer-Verlag.