A global relaxation labeling approach to coreference resolution

Abstract

This paper presents a constraint-based graph partitioning approach to coreference resolution solved by relaxation labeling. The approach combines the strengths of groupwise classifiers and chain formation methods in one global method. Experiments show that our approach significantly outperforms systems based on separate classification and chain formation steps, and that it achieves the best results in the state of the art for the same dataset and metrics.