Unsupervised Event Coreference for Abstract Words
Dheeraj Rajagopal, Eduard H. Hovy, Teruko Mitamura · 2016
We introduce a novel approach for resolving coreference when the trigger word refers to multiple (sometimes non-contiguous) clauses.Our approach is completely unsupervised, and our experiments show that Neural Network models perform much better (about 20% more accurate) than traditional feature-rich baseline models.We also present a new dataset for Biomedical Language Processing which, with only about 25% of the original corpus vocabulary, still captures the essential distributional semantics of the corpus.