Based on Consistency Prototype Learning and Multi-knowledge Distillation for Few-Shot Continual Relation Extraction

Few-Shot Continuous Relation Extraction (FSCRE) has been plagued by issues such as overfitting, catastrophic forgetting, and feature distortion. This paper proposes a few-shot incremental relation extraction model based on consistency prototype learning and multi-layer knowledge distillation. The model consists of data augmentation, consistency prototype learning, and dual-layer knowledge distillation. First, data augmentation is applied to the training data to alleviate common overfitting problems in few-shot scenarios. Second, consistency prototype learning is utilized to regularize the embedding space and address feature distortion. Additionally, a dual-teacher model for knowledge distillation is designed to prevent catastrophic forgetting of previously acquired knowledge. Experiments conducted on the TACRED dataset show that after completing all tasks under the 5-way-5-shot setting, accuracy improved by 2.22%. Our model significantly outperforms other baseline methods, validating the effectiveness of our proposed approach.