KEPA-CRF: Knowledge expansion prototypical amortized conditional random field for few-shot event detection

Abstract

Event Detection (ED) has long struggled with the ambiguous definition of event categories, making it challenging to accurately classify events. Previous endeavors aimed to tackle this problem by constructing prototypes for specific event categories. However, they overlooked potential correlations among instances of distinct event categories, resulting in trigger misclassifications across event types. In this research, we introduce KEPA-CRF to train enhanced event prototypes and address the issue of limited samples in few-shot event detection. By integrating external knowledge from the Glove knowledge base into the model training process, we augment synonymous examples, mitigating the problem of insufficient samples in few-shot event detection. Additionally, through prototype adversarial training, we contrast prototypes of different event categories to augment the representational capabilities of prototype vectors. Experimental results showcase that our approach attains superior performance on the benchmark dataset FewEvent, surpassing comparative models with reduced training time.