Hierarchical Mutual Prompt for Chinese Few-Shot Event Detection

Few-shot event detection (FSED) refers to the task of detecting and classifying events with limited labeled training examples. Existing works mainly focus on utilizing meta-learning to overcome data scarcity by learning and selecting models through identification of triggers and event classification prompts. However, they neglect the similarities between prompts. In addition, most existing methods are designed for English datasets without considering the language differences and structural complexity in Chinese. Therefore, we propose HiPrompt, a hierarchical mutual prompt framework to alleviate the classification bottleneck in Chinese FSED. In detail, we first design an External Dictionary Injection (EDI) module to enrich semantic prompts for Chinese trigger words. Then, we propose a Hierarchical Mutual Prompt (HMP) module that captures underlying semantic relationships between triggers and event types by merging shared prompt components to improve contextual understanding. After that, we propose a hierarchical prototypical network (HPN) module to enhances the generalization ability in representing few-shot event types. Experiments demonstrate that compared with a series of baselines, HiPrompt achieves significant advantages on Chinese FSED datasets and exhibits stronger robustness given extremely scarce training instances.