Employing Gated Mechanism to Incorporate Symbolic Features into Chinese Event Coreference Resolution

Event mentions are sparsely distributed in unstructured text where single event mention without coreferential relationship account for a large proportion. And most of the current studies focus on English events, whereas Chinese has the characteristics of pro-drop and zero entity coreference, making it more challenging to solve Chinese event coreference. In order to solve within-document Chinese event coreference, a new Gated Mechanism Neural Network (GMNN) based on the basic characteristics of events is proposed. First, the pre-trained language model BERT and a feedforward neural network are introduced to represent a small number of necessary basic features(event triggers and event arguments), and the vector representation of symbolic features (event subtypes and event basic attributes) is obtained by trainable embedding matrices; Then, the gated mechanism is used to filter the noise from the symbolic features, extract the useful information in the specific context, and form the event mention-pair representation together with the basic features; Last, a feedforward neural network is applied to calculate the coreference score of the mention-pair representation and output the coreference event cluster. The experiments on ACE2005 Chinese dataset show that the performance of GMNN is improved by 8.2% compared with the baseline.