SemEAGAT: A Novel Approach by Incorporating Semantic Dependency Graph in Event Detection
Yuqi Xi, Jiamin Lu, Tingting Hang, Yunfei Zhang, Zhongyi Wang, Jun Hong Feng · Journal of Physics Conference Series · 2022
Abstract Event detection (ED) is a task that requires capturing deep semantic information in text to correctly identify specific types of events. Semantic dependency graph aims to recover sentence-internal predicate-argument relationships. However, recent ED researches often use syntactic dependency tree in GNNs, while we believe that the semantic dependency graph can further improve the ED task’s effectiveness since it has already been used in many other NLP tasks such as relation extraction, machine translation and abstractive summarization to provide effective semantic information, which is exactly required for event detection. In this paper, we propose a novel semantic dependency edges aware graph attention network (SemEAGAT). It incorporates the semantic dependency graph with an additional multi-head attention in an edge-aware way. Experiments on ACE2005 show our proposed method can achieve better effectiveness by comparing with the state-of-the-art methods.