MIIGraph: Multi-granularity Information Integration Graph for Document-Level Event Extraction

Document-level Event Extraction (DEE) involves extracting event-related structural information, such as event types and event arguments, from a document containing multiple sentences. This task presents challenges, including argument scattering, multiple events, and role overlap, compared to Sentence-level Event Extraction (SEE). Existing works construct heterogeneous graphs for DEE to capture the interactions between entities and sentences. However, they neglect the importance of the global theme information and the interaction information between entities, sentences, and global theme information. To address this gap, we propose the Multi-granularity Information Integration Graph (MIIGraph) framework for DEE. This model aims to capture the interaction of multi-granularity information such as entities, sentences, and global theme of a document for DEE. Specifically, we first obtain the global theme representation of the document through contrastive learning. Then, we construct a heterogeneous graph to capture the complex interactions between entities, sentences, and global theme. Finally, we conducted extensive experiments to evaluate MIIGraph on two widely used DEE benchmarks. The results show that MIIGraph significantly improves the performance of DEE compared to existing methods.