Named entity recognition method based on boundary-aware attention mechanism and generative adversarial network

Named Entity Recognition (NER) aims to locate and identify entities with specific meaning in text. The NER problem can usually be regarded as a type of sequence labeling problem. The key to solving this type of problem lies in determining the boundaries and categories of entities. However, due to the fuzziness of entity boundaries and limitations at the labeling level, most existing NER models introduce vocabulary information loss and The problem of entity boundary recognition error. To this end, a named entity recognition method based on the boundary-aware attention mechanism is proposed. By introducing pointer annotation to construct the boundary position vector, a sequence annotation layer that fuses the boundary position information is established to fully exploit the boundary characteristics of the entity. On the basis of integrating the boundary position vector, the lattice structure of the text information is converted into a planar structure composed of spans, a dynamic position encoding strategy is designed based on pointer annotation, and then the semantics of the label of the entity annotation and the entity boundary position are learned based on the generative adversarial network Similarity, and improve entity recognition performance by introducing information of entity boundary pointers in the weight calculation of the attention mechanism. Experimental results on weibo and Chinese-Literature-NER data sets show that the proposed method has obvious advantages in accuracy and F1 index compared with the baseline method, verifying the effectiveness of the attention mechanism based on boundary awareness in named entity recognition.