Unleashing Pre-trained Masked Language Model Knowledge for Label Signal Guided Event Detection

Mengnan Xiao, Ruifang He, Junwei Zhang, Jinsong Ma, Haodong Zhao

Abstract: Event detection (ED) aims to recognize triggers and their types in sentences. Previous work employs distantly supervised methods or pre-trained language models to generate sentences containing events to alleviate data scarcity. Further, determining the spans and types of triggers is complex and may have deviations. In this paper, we propose to unleash Pre-trained Masked Language Model (PMLM) knowledge for label signal guided ED by a novel trigger augmentation. We directly generate triggers by leveraging the rich knowledge of PMLM through masking triggers. However, these newly replaced triggers may not correspond to the label of the masked trigger. To control such trigger augmentation noises, we design a label signal guided classification mechanism with event type-subtype guidance. To ensure the quality of generated triggers, a semantic consistency mechanism is introduced. Experimental results on the ACE2005 and FewEvent show the effectiveness of our proposed approach.