Research on medical text named entity recognition based on Two-stage approach

At present, more medical Named-entity recognition work focuses on extracting continuous entities, ignoring overlapping entities and discontinuous entities. Due to the fact that sequence annotation frameworks (BIO, etc.) suitable for continuous entities are not well suited for the recognition of discontinuous and overlapping entities, these structured entities are inherently difficult to recognize. Therefore, based on the two-stage approach of decomposing entity recognition into two subtasks: segment generation and segment combination, a new annotation framework is considered, and character features are incorporated to enrich the features of discontinuous and overlapping entities in the data. In terms of model construction, BERT input vectors are used to replace Word2Vec pretraining model inputs, and relative position self-attention layers are added to enhance the model's data processing ability. The experimental structure shows that compared to mainstream model frameworks, this framework can better solve the recognition problems of discontinuous and overlapping entities in medical texts.