Link Type Based Pre-Cluster Pair Model for Coreference Resolution.

This paper presents our participation in the CoNLL-2011 shared task, Modeling Unrestricted Coreference in OntoNotes. Coreference resolution, as a difficult and challenging problem in NLP, has attracted a lot of attention in the research community for a long time. Its objective is to determine whether two mentions in a piece of text refer to the same entity. In our system, we implement mention detection and coreference resolution seperately. For mention detection, a simple classification based method combined with several effective features is developed. For coreference resolution, we propose a link type based pre-cluster pair model. In this model, pre-clustering of all the mentions in a single document is first performed. Then for different link types, different classification models are trained to determine wheter two pre-clusters refer to the same entity. The final clustering results are generated by closest-first clustering method. Official test results for closed track reveal that our method gives a MUC F-score of 59.95%, a B-cubed F-score of 63.23%, and a CEAF F-score of 35.96% on development dataset. When using gold standard mention boundaries, we achieve MUC F-score of 55.48%, B-cubed F-score of 61.29%, and CEAF F-score of 32.53%.