Making Latent SVMstruct Practical for Coreference Resolution

The recent work on coreference resolution has shown a renewed interest in the structured perceptron model, which seems to achieve the state of the art in this field. Interestingly, while SVMs are known to generally provide higher accu- racy than a perceptron, according to pre- vious work and theoretical findings, no re- cent paper currently describes the use of SVMstruct for coreference resolution. In this paper, we address this question by solving some technical problems at both theoretical and algorithmic level enabling the use of SVMs for coreference resolu- tion and other similar structured output tasks (e.g., based on clustering).