Continual few-shot named entity recognition against catastrophic forgetting and overfitting

Named Entity Recognition (NER), a fundamental task in information extraction, identifies and classifies pre-defined entity types in text. In dynamic real-world scenarios, NER systems must continuously learn new entity types from only a few labeled examples, presenting challenges of catastrophic forgetting and few-shot overfitting. To address these issues, this paper proposes the Prompt-guided Memory-Knowledge augmentation with Contrastive and knowledge Distillation (PMKCD) framework. Our approach uses label prompting to explicitly model category semantics, enhancing discrimination in low-resource settings. It further introduces a novel data augmentation strategy that combines a dynamic memory set with knowledge-guided, dual-granularity replacement to generate high-quality synthetic samples. This facilitates balanced knowledge integration across old and new classes. The framework employs contrastive distillation to jointly optimize model stability and plasticity. Extensive experiments on three public benchmarks show that PMKCD consistently outperforms existing methods in continual few-shot NER. The results show an average relative improvement of 14.75% in Micro-F1 and 8.69% in Macro-F1 across all incremental stages, along with significant gains in recognition accuracy, forgetting mitigation, and generalization capability.