A Neural Span-Based Continual Named Entity Recognition Model

Named Entity Recognition (NER) models capable of Continual Learning (CL) are realistically valuable in areas where entity types continuously increase (e.g., personal assistants). Meanwhile the learning paradigm NER advances to new patterns such as span-based methods. However, its potential CL has not been fully explored. In this paper, we propose SpanKL, a simple yet effective Span-based model with Knowledge distillation (KD) preserve memories and multi-Label prediction prevent conflicts CL-NER. Unlike prior sequence labeling approaches, inherently independent modeling span level designed coherent optimization on SpanKL promotes at each incremental step mitigates forgetting. Experiments synthetic datasets derived from OntoNotes Few-NERD show that significantly outperforms previous SoTA many aspects, obtains smallest gap upper bound revealing high practiced value. The code is available https://github.com/Qznan/SpanKL.