Context-Aware Prompt for Generation-based Event Argument Extraction with Diffusion Models

Abstract

Event argument extraction (EAE) has attracted increasing attention via generation-based methods. However, most existing works tend to independently extract arguments for each role, ignoring the correlation between different arguments, especially in long contexts. Motivated by these observations and the high-quality generation results of recent diffusion models, we propose an effective model called PGAD (Context-Aware Prompt for Generation-based EAE with Diffusion models) for both sentence-level and document-level EAE. In PGAD, a text diffusion model is designed to generate diverse context-aware prompt representations in conjunction with a series of random Gaussian noise. Firstly, cross-attention is employed between the designed prompt and input context within the text diffusion model in order to generate the context-aware prompt. Through this interaction, the context-aware prompt is able to capture multiple role-specific argument span queriers. Secondly, the context-aware prompt is aligned with the context to generate event arguments by joint optimization. Extensive experiments on three publicly available EAE datasets demonstrate the superiority of our proposed PGAD model over existing approaches.