Self-distilled Named Entity Recognition Based on Boundary Detection and Biaffine Attention

Named Entity Recognition (NER) is an important down-streaming task in natural language processing. Span-based methods are applicable to both flat and nested entities. However, they lack explicit boundary supervision. To address this issue, we propose a multi-task and self-distilled model which combines biaffine span classification and entity boundary detection tasks. Firstly, the boundary detection and biaffine span classification models are jointly trained under a multi-task learning framework to address the problem of lacking supervision of boundaries. Then, self-distillation technique is applied on the model to reassign entity probabilities from annotated spans to surrounding spans and more entity types, further improving the accuracy of NER by soft labels that contain richer knowledge. Experiments were based on a high-density entity text dataset of the commodity titles from an e-commerce company. Finally, the experimental results show that our model exhibited a better F1 score than the existing common models.