Employing Gated Attention and Multi-similarities to Resolve Document-level Chinese Event Coreference

Abstract

Event coreference resolution is a challenging task. To address the issues of the influence on event-independent information in event mentions and the flexible and diverse sentence structure in Chinese language, this paper introduces a GANN (Gated Attention Neural Networks) model to document-level Chinese event coreference resolution. GANN introduces a gated attention mechanism to select eventrelated information from event mentions and then filter noisy information. Moreover, GANN not only uses a single Cosine distance to calculate the linear distance between two event mentions, but also introduces multi-mechanisms, i.e., Bilinear distance and Single Layer Network, to further calculate the linear and nonlinear distances. The experimental results on the ACE 2005 Chinese corpus illustrate that our model GANN outperforms the state-of-the-art baselines.