Few-Shot Event Extraction with a Dialogue-based Large Language Model

Event extraction is critical in various fields, including knowledge graph construction, public opinion monitoring, and situational awareness. Existing methods rely on supervised learning, which depends heavily on the scale and quality of the dataset, while the model’s generalization ability remains weak. As a result, their application in real-world scenarios is significantly constrained. To address this issue, we propose a few-shot event extraction method that harnesses large language models’ reading comprehension and event extraction capabilities. Our approach decomposes the task into multiple stages within a dialog-based framework. Experimental results show that our method outperforms most few-shot techniques, achieving a 4.1% performance improvement over ChatIE.