Document-level Multi-event Extraction Model Based on Argument Correlation and Graph Neural Network

LIAO Tao; HAO Juanjuan

doi:10.13501/j.cnki.42-1908/n.2026.03.019

Abstract：To address the issues of missing global semantic correlations between events and argument roles and insufficient utilization of document information in document-level multi-event extraction,a document-level multi-event extraction model based on argument correlation and graph neural network(DEEACG)was proposed.Firstly,entities were obtained using bidirectional encoder representations from Transformers(BERT)module,and an entity co-occurrence prediction task was introduced to enhance semantic associations among entities.Secondly,learnable event proxy nodes were incorporated to construct a heterogeneous graph consisting of entities,contextual information,and proxy nodes.Through the graph neural network with feature-wise linear modulation(GNN-FiLM)and multi-head self-attention mechanisms,global interactions and semantic fusion among multiple events were achieved.Thirdly,event type detection was performed using a multi-layer perceptron.Finally,a dual-projection space was constructed to model argument correlations.The Bron-Kerbosch algorithm was applied to extract maximal cliques from the graph as candidate argument combinations,and multi-head attention was utilized for argument role classification.The results demonstrated that DEEACG model was evaluated on the Chinese financial announcements(ChFinAnn)dataset,where its performance in multi-event extraction tasks showed significant improvement.Compared with the relation-enabled document-level event extraction(ReDEE)model,the mean value of F1-score was increased by 2.1 percentage points.This study confirmed that DEEACG model could effectively capture semantic associations among multiple events and was suitable for document-level multi-event extraction tasks.