A Graph Convolution Network with a POS-aware Filter and Context Enhancement Mechanism for Event Detection

Event detection (ED) is a task that focuses on identifying event instances in texts. Many previous studies have demonstrated the effectiveness of syntactic dependency trees and graph convolution networks (GCN) in ED tasks. However, such methods haven't utilized the correlation between part-of-speech (POS) and keyword distribution to filter the noise from graph convolution. In addition, owning to the over-smoothing problem of GCN, their abilities in contextual understanding also need to be improved. In this paper, we propose a novel graph convolution network with a POS-aware filter and context enhancement mechanism (GCN-PFCE). Specifically, a gating unit controlled by POS, which can learn the correlation between POS and keyword distribution, is added after each graph convolution layer. Besides, a parallel structure between BERT and GCN is implemented to enhance the context understanding ability of GCN-based methods in a better way. The proposed model achieves significant improvement over competitive baseline methods on ACE2005 dataset. Additionally, the code of this paper is released as open-source at https://github.com/Jiansheng-Chen/GCN-PFCE.