Coreference Clustering using Column Generation
Jan De Belder, Marie‐Francine Moens · Lirias · 2012
In this paper we describe a novel way of generating an optimal clustering for coreference resolution. Where usually heuristics are used to generate a document-level clustering, based on the output of local pairwise classifiers, we propose a method that calculates an exact solution. We cast the clustering problem as an Integer Linear Programming (ILP) problem, and solve this by using a column generation approach. Column generation is very suitable for ILP problems with a large amount of variables and few constraints, by exploiting structural information. Building on a state of the art framework for coreference resolution, we implement several strategies for clustering. We demonstrate a significant speedup in time compared to state-of-the-art approaches of solving the clustering problem with ILP, while maintaining transitivity of the coreference relation. Empirical evidence suggests a linear time complexity, compared to a cubic complexity of other methods.