EKE: External Knowledge-Enhanced Prompt for Chinese Few-Shot Event Extraction

Shuxiang Hou, Yurong Qian, Jiaying Chen, Hongyong Leng

Abstract: Chinese few-shot event extraction faces significant challenges due to the limited availability of labeled data and the inherent complexity of Chinese linguistic features, which often lead to ambiguity in event boundaries and semantics. To tackle these issues, we first annotate a dataset specifically for Chinese few-shot event extraction. Then, we propose EKE (External Knowledge-Enhanced Prompt), a novel prompt-based framework designed for this task. EKE consists of three key components: external knowledge prompts, event span comparison detection, and event argument hierarchical prompts. Notably, the external knowledge prompt incorporates lexical information from Chinese dictionaries and syntactic dependencies to enrich semantic representations, thereby improving trigger identification. A contrastive learning strategy is employed during event detection to model fine-grained span boundaries and enhance prototype representation. For argument extraction, we introduce a hierarchical prompting mechanism along with a bipartite matching loss to optimize argument span assignment while mitigating prototype noise introduced by external knowledge. Experimental results on benchmark datasets show that EKE significantly outperforms baseline methods, demonstrating its effectiveness and generalizability in Chinese few-shot event extraction.

External IDs:doi:10.1007/978-981-96-9994-0_38