WCSE

Event Detection via Graph-Based Multi-hop Neighbors’ Information Fusion

Abstract— Event detection (ED) proves to be a crucial subtask of event extraction. There is a close connection between event trigger word and its related neighboring words in dependency parse tree, neighbors’ information has been widely used in event detection task. However, only neighbors’ information with a direct arc (i.e., only one hop) to the trigger candidate in the dependency parse tree is used in many existing graph convolutional network (GCN)-based methods. As a result, the multi-hop neighbors’ information is not fully utilized, so that it is difficult to further improve the performance of ED. Therefore, this paper proposes an ED model of graph convolutional network based on dependency parse tree. This model introduces graph attention network (GAT) to learn the multi-hop neighbors’ information of each adjacent node in the syntactic graph, and uses a multi-label attention fusion mechanism to fuse the extracted multi-hop semantic information and reduce the complexity of the model. The results of experiments on the ACE2005 dataset show that compared with other methods based on graph convolutional neural networks, the F1-socre of this method reaches 75.6%, and the experimental results reflect the effectiveness of the method proposed in this paper.

Index Terms— event detection, dependency parse tree, Graph convolution network, Graph attention network, attention fusion mechanism.

Chuan Li
Xi'an University of Posts and Telecommunications
Guoqiang Tian
Xi'an University of Posts and Telecommunications
Xintong Sun
Xi'an University of Posts and Telecommunications
Fang Yang
Xi'an University of Posts and Telecommunications

[Download]

Cite: Zhenchong Mo, Lin Gong, Ziyao Huang, Mingren Zhu, Xie Jian, Junde Lan, Shicheng Zhang, " Event Detection via Graph-Based Multi-hop Neighbors' Information Fusion, " WCSE 2022 Spring Event: 2022 9th International Conference on Industrial Engineering and Applications, pp. 1558-1567, Sanya, China, April 15-18, 2022.