Attention-Based Graph Convolution Networks for Event Detection
Zhu Hanqing, Kaiming Xiao, Lizhen Ou, Mao Wang, Lihua Liu, Huang Hongbin · 2021
Event detection is an important subtask in the field of event extraction. Most of the current state-of-the-art(SOTA) methods introduce rich external information to improve the performance of detection. The graph convolution networks (GCNs) method has been proved to be an effective way to enhance the semantic representation of each word, in which an adjacency matrix is constructed by analyzing the dependencies among the words in the corpus. Syntactic structure and typed dependency label information has shown its positive effects on performance enhancement of event detection. However, few works consider multi-order distance and edge representation update of dependency graph at the same time. Therefore, we design a new way to combine edge representation update method based on attention weight and multi-order distance edge labels to enhance the semantic impact of long range dependency. The individual and composite effects of these two enhancements have been analyzed. Experiments on the ACE2005 dataset show a performance improvement of 0.4% over SOTA GCNs baseline method.