Medical Record Entity And Relation Extraction Technology Integrating Multiple Features And Attention Mechanism

Intelligent medical service is a healthcare service model based on artificial intelligence and internet technique,which can provide patients with more intelligent and convenient medical services,thus improving medical efficiency and medical quality.How to effectively extract key information from patients’ medical records and transform unstructured medical information into structured data has long been a pressing problem in the field of intelligent medical service.To address this challenge,efficient and intelligent methods are used to extract medical entities and relationships from medical records is a key research task in the field of intelligent medical service.However,there is currently limited research on Chinese medical entity recognition and relationship extraction tasks,and many challenges are faced:(1)existing entity recognition research methods don’t fully consider the language characteristics of Chinese medical records,and utilize Chinese medical features insufficiently;(2)existing entity and relationship extraction research methods don’t fully consider the interaction between entities and relationships,and the information between the two sub-tasks is not fully utilized.To address these challenges,the main works of this thesis are as follows:The thesis proposes a Chinese medical record medical entity recognition model based on the dual-branch TENER.Firstly,in the process of encoding Chinese medical text using the TENER pre-training model,the Char-Entity-Transformer is used to integrate medical entity dictionary information into the model,helping TENER obtain more feature information of Chinese medical terminology.Secondly,a divide-andconquer strategy is proposed to divide entity recognition into two tasks: entity boundary recognition and entity type recognition.The entity boundary recognition task uses CRF to obtain entity boundary information;The entity type recognition task determines the entity’s category using A-Softmax.Finally,in order to improve the model’s classification ability for entities,the radical features which extracted from Chinese medical text using convolutional neural networks are incorporated into the entity type recognition task.Experimental results on four Chinese medical record datasets validate the effectiveness of the model.The thesis proposes a Cross-TENER model with fusion attention-guided graph convolution to jointly extract Chinese medical entities and relations in medical records.This model first uses attention-guided graph convolutional networks to learn the dependency structure of the sentence,selectively focusing on relevant information useful for relation extraction.Then,the Star-Transformer’s self-attention mechanism is used to introduce entity boundary information to enhance entity boundaries and better capture the positional relationships between different words in the sentence,further enhancing the model’s expressive power.Finally,Cross-TENER calculates the correlation of different medical entity relationships,capturing the connection between different relationships,the fine-grained semantic features in words and the boundary information of medical entities,reducing the noise brought by irrelevant relationships in entity recognition.Experimental results show that compared to other baseline models,the model proposed in this this thesis performs better in three Chinese medical datasets and helps improve the performance of joint extraction of Chinese medical entities and relations.