A Comprehensive Survey of Continual Learning Techniques in NER: Challenges and Opportunities

Continual learning is a paradigm that studies the problem of learning from a stream of information without forgetting previously learned knowledge. This literature review explores the challenges and opportunities of continual learning for Named Entity Recognition (NER). A central concern with continual learning is catastrophic forgetting, where training a neural network on new information leads to degradation in performance on previously learned information. Traditional reannotation and model retraining methods prove costly and time-consuming, prompting the need for better strategies. We introduce previous work reviewing the area of continual learning in Section 2. We explore the evolution of continual learning in Section 3.1, categorizing works into Replay, Regularization, and Parameter Isolation approaches. We address evaluation metrics in Section 3.2, detailing how to measure the effectiveness of continual learning. Section 4 compares continual learning methods specific to NER in the three main categories defined in Section 3.1. Section 5 presents an evaluation of the reviewed NER systems. Finally, Section 6 discusses challenges, emerging trends, and future directions, providing a comprehensive overview of the application areas of continual learning for NER. This review contributes to a deeper understanding of continual learning’s role in advancing NER and identifies avenues for further research and innovation.