An improved GPT2-based joint event extraction method with position expansion and knowledge augmentation.

With the rapid development of the Internet and social media, massive amounts of unstructured data have emerged, making event extraction increasingly important for information retrieval. In the financial domain, challenges such as long texts, redundant content, and complex structures hinder extraction tasks. To address this, we propose PosEKE-GPT2, an improved GPT2-based model that reformulates event extraction as a text generation task. The model jointly identifies event types, triggers, and arguments using structured canonical text and a sub-task extraction strategy to reduce error propagation. An expanded positional encoding mechanism enhances event representation in long texts. Furthermore, we introduce a knowledge augmentation module that dynamically selects and integrates external knowledge via prompt mechanisms and attention-based embedding optimization. Experiments on the DuEE-Fin dataset show that PosEKE-GPT2 achieves an average F1-score of 90.61, while on the FewFC dataset it reaches an average F1-score of 88.85, both outperforming baseline models. Ablation studies verify the effectiveness of the positional encoding and knowledge augmentation modules, demonstrating the model's robustness and suitability for financial event extraction across different datasets.