Nested Named Entity Recognition based on Span and Efficient Global Pointer

Named Entity Recognition (NER) method based on deep learning has made some achievements in the field of natural language processing (NLP). However, with the increasing complexity of text information structure, NER methods based on sequence tagging make it difficult to solve the problem of entity nesting and face the problem of negative samples in the process of model training. In this paper, an efficient global pointer (SEGP) method based on span is designed and try to solve the class imbalance. We construct a global entity matrix by combining the Multi-Head Attention mechanism and rotating position encoding to make full use of boundary information. We also propose the sample weight adjustment loss to solve the class imbalance problem. Finally, experiments on the ACE 2005 corpus and GENIA corpus show that the precision is improved by 0.72% and 2.05% respectively. The model designed in this paper can effectively identify nested entities.