A Study of Chinese Medicine Entity Recognition Method by Fusing Multi-Features and Pointer Networks.

The recognition of named entities in Traditional Chinese medicine (TCM) is a difficult task in the field of medical information extraction, which often contains a large number of domain nouns and specialized terms with high semantic complexity and unclear entity boundaries and multiple meanings among some entities. In order to effectively solve the recognition problem of named entities in TCM, and to address the phenomenon of underutilized semantic information in entity recognition tasks, an entity recognition method incorporating Chinese character multi-features and SPAN pointer networks is proposed to obtain character feature vectors of data using the powerful characterization information of the pre-training model BERT, connect the character vectors with lexical and radical feature embeddings, and obtain the long-range textual context through BiGRU and Attention layer to obtain the contextual information of long-range text, and finally use SPAN pointer network to achieve the start boundary determination of the entity and complete the extraction of the entity. In addition, adversarial training and focal loss are added to reduce the risk of overfitting and enhance the generalization ability and robustness of the model. The final experiments show that this method has superior performance in dealing with the named entity recognition problem of TCM.