TGIN: Document-level event extraction with two-phase graph inference network

Document-level event extraction aims to extract event records from a whole document that contain numerous entities scattered across multiple sentences. Efficiently modeling the interactions among these entities is crucial. However, previous methods suffer from two main shortcomings. Firstly, they tend to implicitly model key information, which can result in representations with higher levels of noise. Secondly, they excessively consider irrelevant entities, thereby reducing extraction efficiency and precision. To address these issues, we propose a novel Two-phase Graph Inference Network (TGIN) approach for extracting document-level events. In the first phase, TGIN constructs a heterogeneous document-level graph to capture complex interactions among nodes of different granularity, enabling the acquisition of document-aware features. Subsequently, a dedicated module is developed to extract relevant entity pairs within the same event record. This module utilizes a key information aggregator with an attention mechanism to explicitly aggregate key sentences for entity pairs. In the second phase, the entity links predicted in the first phase serve as prior information to construct the entity-level graph, which focuses on modeling interactions between entity pairs that potentially share the same event link, effectively reducing error propagation. Experimental results on the publicly available document-level event extraction dataset ChFinAnn demonstrate the superiority of our framework over most existing models.