A Document Level Event Extraction Method Without Triggers

Event extraction aims to extract event information from massive unstructured texts, and has broad application prospects. Unlike sentence level event extraction, document level event extraction faces higher complexity and difficulties in missing trigger words. In recent years, document level event extraction models based on graph structures have gradually emerged, but there are still shortcomings such as high computational and spatial costs, and poor graph structure prediction ability. This article proposes an improved method from PTPCG to address the above issues. The model has stronger encoding ability and introduces Biaffine in adjacency matrix prediction to improve the performance. Ultimately, the performance of event extraction tasks on the Duee-fin dataset was improved, achieving the best recall and F1 value compared to the baseline model.