Joint Extraction Of Entity-Relation-Attribute For Application Security In Evidence-Based Medicine

With the development of medical research,the number of literature in related fields grows rapidly.Evidence-based medicine researchers need to read a large number of literature to obtain the latest research results.The evidence-based medicine application security is directly related to medical security,and the timeliness,correctness and integrity of medical evidence acquisition directly affect the evidence-based medicine application security.Although experts in the field will summarize systematic reviews for reference,summarizing reviews takes a lot of time and energy,so automatic extraction of medical evidence has become an important topic in promoting evidence-based medical practice.The existing research on medical evidence extraction is usually sentence-level summary extraction or identification of a few important attributes,and there are very few related studies on fine-grained medical evidence extraction.Medical evidence summary can speed up the acquisition of medical evidence,but it is unstructured information,which cannot directly construct structured medical evidence.And identifying a few important attributes is not enough to comprehensively describe medical evidence,which is difficult to ensure the evidence-based medicine application security.Based on the above background and the existing Aceso evidence extraction system in the laboratory,this thesis further studies the joint extraction of entity-relation-attribute method for application security in evidence-based medicine.The method is based on the PICO framework.On the basis of extracting medical evidence summary,the fine-grained entity relation and attribute in medical evidence summary are extracted as accurately and completely as possible to ensure the application security of evidence-based medicine.The main research contents are as follows:(1)To solve the problems of expensive manual annotation cost and low domain generalization faced by medical evidence knowledge acquisition,this thesis proposes an evidence-based medicine summary extraction method based on deep learning(Aceso-SEME).This method automatically extracts medical evidence-related sentences from medical literature to construct medical evidence summary.Different from Aceso’s supervised training,we uses a deep learning model for text classification.The language model trained on the biomedical corpus is selected for the text vector representation,and a classification model including a bidirectional recurrent neural network and an attention mechanism is designed.After classifying,the sentences related to evidence-based medicine will finally to be output to form a medical evidence summary,and the effectiveness of the method is proved by experiments.Method reduces the cost of manual labeling,enhances the generalization of evidence extraction in different medical fields,improves the efficiency of medical evidence acquisition and the extraction quality of medical evidence summary,and ensures the application security of evidence-based medicine.(2)To solve the problems that the fine-grained medical evidence extraction effect is poor,the information extraction pays more attention to entity relation but ignores entity attribute,and contains implicit subjects in medical evidence-related sentences,this thesis proposes a joint extraction of fine-grained entity-relation-attribute method in medical evidence(Aceso-Joint).Structural medical evidence is extracted from medical evidence summaries in the form of triples,which are used to build evidence-based medicine knowledge graphs and apply to other downstream tasks.The method using the deep learning method,the language model trained on the biomedical corpus was selected to represent the text vector,and the joint extraction model was used to extract the entity relation and entity attribute triples to obtain structured medical evidence.The effectiveness of the method is proved by experiments.The method takes into account the extraction of entity relation and entity attribute,and solves the problem of implicit subjects in medical evidence-related sentences,which improves the quality and integrity of fine-grained medical evidence extraction as a whole,and ensures the application security of evidence-based medicine.(3)Design and implement the medical evidence extraction visualization system-Aceso2.0.On the basis of extracting medical evidence summary,the system simultaneously extracts fine-grained medical evidence triples.System integrates the Aceso-SEME method and the Aceso-Joint method.The main functions include uploading documents,extracting medical evidence summary and fine-grained medical evidence,visualizing medical evidence and downloading results.