Entity-Aware Graph Convolution Networks for Event Detection

Abstract

The existing event detection models based on graph convolutional networks only consider the syntactic structure from the dependency tree of the sentence, and ignore the importance of the entity to the event in the syntactic structure, which is used for event detection task. In this paper, we propose a Entity-Aware Convolutional Networks (EAGCN) which exploits adding entity information directly to the syntactic structure by dynamically modifying the dependency graph of the sentence in the convolution operation. Besides, we followed EAGCN with a Bi-directional Long-Short Term Memory to import sequence information into structure information, which is an indispensable part of the model. The sufficient experimental results show that our model achieves the best F1 score for the event detection task on the ACE 2005 dataset.