Contextualized Hybrid Prompt-Tuning for Generation-Based Event Extraction

Yuzhen Zhong, Tong Xu, Pengfei Luo

Abstract: Recent years have witnessed the wide attention of event extraction task, which highly benefits various downstream applications. Traditionally, prior arts attempt to model this task in the perspective of sequence generation problem with discrete prompts. However, these methods typically construct templates for each event type without considering unique instance characteristics, which may lead to the suboptimal performance and unwanted noise. To address this issue, in this paper, we propose a novel solution for event extraction based on structure generation with contextualized hybrid prompt tuning, called CHPT-EE, in terms of both discrete and continuous prompts. Specifically, CHPT-EE unifies the encoding of event mentions from various event types via a structured event extraction language. Along this line, CHPT-EE could effectively exploit the complementary advantages of different prompts to mitigate the issues encountered by prior methods. Afterwards, for each context, we employ a semantically similar strategy to retrieve answered prompts as discrete prompt DISC-P, and obtain context-aware continuous prompt CONT-P by aggregating specific event type information and contextual semantic information. Experiments on ACE 2005 and CASIE show that CHPT-EE achieved competitive performance compared with state-of-the-art methods.