Efficient event extraction via one-stage dynamic fusion: a novel approach for overlapping and nested events

Event Extraction (EE) constitutes a core task within information extraction., with the main objective of obtaining structured event data from unstructured text. Traditional methods often focus on extracting single events, frequently overlooking the existence of overlapping or nested events. Although some models attempt to extract event triggers and arguments through multiple sequential stages to handle overlapping and nested events, these models are susceptible to error propagation, which reduces the accuracy and reliability of the extraction. To address these challenges, we propose a one-stage dynamic fusion approach for extracting overlapping and nested events. This framework consists of a BERTbased Encoding Layer, a Dynamic Fusion Processing Layer, and a Cascade Global Prediction Layer. The dynamic fusion layer uses cross-normalization, self-normalization, attention mechanisms, and gating mechanisms to integrate the target event types with word representations. The cascade prediction layer leverages the KAN predictor to extract triggers, arguments, and their relationships in a single stage, thereby eliminating error propagation. Extensive of experiments on the FewFC dataset demonstrate that our model attains state-of-the-art (SOTA) results in the field, significantly improving the performance of argument classification. Our contributions include: (1) introducing a dynamic fusion layer that efficiently combines both local and global event information; (2) proposing a one-stage framework that efficiently handles overlapping and nested events; (3) outperforming existing methods on the FewFC dataset.