Event Argument Extraction with Enriched Prompts
Chen Liang
TL;DR
This paper investigates the ceiling and effectiveness of prompt-based event argument extraction (EAE) by varying the information content in prompts. It proposes three prompt-model variants with increasing information density—single-role, multi-role, and multi-event prompts—and an enhanced model using a region (dice) loss to improve training. Empirical results on RAMS across multiple transformers show that intra-event information yields larger gains than inter-event information, and that well-designed prompts plus loss regularization can still improve performance, though large language models can exhibit hallucinations and require careful evaluation. The work provides upper-bound concepts for each prompt class and highlights practical guidance for exploiting intra-event dependencies in EAE tasks.
Abstract
This work aims to delve deeper into prompt-based event argument extraction (EAE) models. We explore the impact of incorporating various types of information into the prompt on model performance, including trigger, other role arguments for the same event, and role arguments across multiple events within the same document. Further, we provide the best possible performance that the prompt-based EAE model can attain and demonstrate such models can be further optimized from the perspective of the training objective. Experiments are carried out on three small language models and two large language models in RAMS.