Event Relation Extraction Using Type-Guided Attentive Graph Convolutional Networks

Ling Zhuang, Po Hu, Weizhong Zhao

Abstract: Event relation extraction is a fundamental task in text mining, which has wide applications in event-centric natural language processing. However, most of the existing approaches can hardly model complicated contexts since they fail to use dependency-type knowledge in texts to assist in identifying implicit clues to event relations, leading to the sub-optimal performance on this task. To this end, we propose a novel type-guided attentive graph convolutional network for event relation extraction. Specifically, given the input text, the event-specific syntactic dependency graph is first constructed, from which both the local and global dependency knowledge related to events are derived. Then, a dependency-type guided attentive graph convolutional network is designed for learning representations of events, in which the local and global dependency information are utilized to effectively aggregate semantic context among texts. Finally, the event relation is predicted based on the representations of event pair and the representation of the whole text, completing the task of event relation extraction. The experimental results on multiple datasets show that our method significantly outperforms the state-of-the-art baselines. Moreover, further analysis reveals that our method can effectively capture complex relationships between long-distance events, improving accordingly the performance of event relation extraction.