Latent tree structure learning for cross-document coreference resolution

[摘要] Cross Document Coreference Resolution (CDCR) is the problem of learning which mentions, coming from several different documents, correspond to the same entity. This thesis approaches the CDCR problem by first turning it into a structure learning problem. A latent tree structure, in which leaves correspond to observed mentions and internal nodes correspond to latent sub-entities, is learned. A greedy clustering heuristic can then be used to select subtrees from the learned tree structure as entities. As with other structure learning problems, it is prudent to envoke Occam;;s razor and perform regularization to obtain the simplest hypothesis. When the state space consists of tree structures, we can impose a bias on the possible structure. Different aspects of tree structure (i.e. number of edges, depth of the leaves, etc.) can be penalized in these models to improve the generalization of thes models. This thesis draws upon these ideas to provide a new model for CDCR. To learn parameters, we implement a parameter estimation algorithm based on existing stochastic gradient-descent based algorithms and show how to further tune regularization parameters. The latent tree structure is then learned using MCMC inference. We show how structural regularization plays a critical role in the inference procedure. Finally, we empirically show that our model out-performs previous work, without using a sophisticated set of features.

[发布机构] Massachusetts Institute of Technology