Few-shot Class Incremental Learning for Event Extraction

In the field of natural language processing, research on incremental learning is relatively scarce compared to computer vision. Most existing studies are limited to tasks such as named entity recognition(NER), automatic question answering and text classification, and have not been applied to event extraction(EE). Considering the characteristics of EE (Continuously emerging new tasks and knowledge; Limited training samples for new tasks in real-life scenarios), our study focuses on a more challenging yet practical problem, i.e., fewshot class incremental learning for EE, which aims to conduct incremental learning under the condition of few samples, enabling it to perform well on new tasks while avoiding the phenomenon of "catastrophic forgetting". To address the fewshot problem, we generate a knowledge-enhanced EE model. and in terms of retaining old knowledge, inspired by knowledge distillation, regularizing constraints on multiple modules. Experimental results demonstrate that our approach effectively addresses the issues and improves the performance of event extraction models.