Multiple event detection via bi-graph attention network with syntactic and semantic features

The complicated inter-dependency among events poses a significant challenge to modeling event correlations in multiple event detection. While graph-based methods outperform sequence-based methods in capturing global features, two limitations may arise. First, redundant syntactic structures and ignored inter-event connections influence explicit relationship modeling. Second, the context-independent nature of syntactic features limits the semantic correlation modeling between events. To alleviate these challenges, we propose Bi-Graph Attention Network for Event Detection (BiGAT-ED), integrating syntactic and semantic features through a multi-perspective graph architecture. Specifically, we propose the Edge-Enhanced Graph Attention Network (EE-GAT) to incorporate syntactic dependency information into the attention mechanism, reducing the impact of redundant connections. We further construct the Event Relation Graph (ERG) to model inter-event connections and encode it with the Node-Aware Graph Attention Network (NA-GAT), which leverages event label knowledge within its attention mechanism for enhanced semantic correlation modeling. Finally, BiGAT-ED integrates event correlations from EE-GAT and NA-GAT through an attention fusion mechanism, improving performance for multiple event identification. Experimental results on cross-domain datasets demonstrate that BiGAT-ED consistently surpasses existing competitive models and leading LLMs.