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

Area of Science:

• Natural Language Processing
• Machine Learning
• Information Extraction

Background:

• Named Entity Recognition (NER) is crucial for information extraction but faces challenges in dynamic environments.
• Continual learning in NER requires systems to adapt to new entity types with limited data, risking catastrophic forgetting and overfitting.

Purpose of the Study:

• To propose a novel framework, Prompt-guided Memory-Knowledge augmentation with Contrastive and knowledge Distillation (PMKCD), for continual few-shot NER.
• To enhance the ability of NER systems to learn new entity types without forgetting previously acquired knowledge.

Main Methods:

• Utilized label prompting to model category semantics and improve discrimination in low-resource scenarios.
• Implemented a data augmentation strategy combining a dynamic memory set with knowledge-guided replacement for synthetic sample generation.
• Employed contrastive distillation to balance model stability and plasticity during continual learning.

Main Results:

• PMKCD consistently outperformed existing methods in continual few-shot NER across three benchmarks.
• Achieved an average relative improvement of 14.75% in Micro-F1 and 8.69% in Macro-F1.
• Demonstrated significant gains in recognition accuracy, forgetting mitigation, and generalization capability.

Conclusions:

• The PMKCD framework effectively addresses the challenges of catastrophic forgetting and few-shot overfitting in continual few-shot NER.
• The proposed methods enable robust and adaptable NER systems for dynamic real-world applications.