Research And Implementation On Medical Literature Abstracts Extraction Methods Based On Knowledge Graph

With the sharply increasing of medical literature,people who take up with Evidence-Based Medicine need read lots of medical literature to obtain the latest research results.It is is timeconsuming and laborious for relying on experts alone to manually summarize the literature evidence,so automatically extracting structured evidence from the medical literature as an abstract is essential to accelerate the practice of evidence-based medicine.The existing medical literature abstract text-driven methods ignore the use of medicine knowledge,resulting in the following problems:(1)Multi-source of medicine knowledge: rich medicine knowledge is stored in multiple vary medical knowledge graphs of different fields.How to use different medicine knowledge to improve the quality of medicine evidence abstract is an urgent problem to be solved.To address this problem,this thesis proposes a method of representation learning about acrossing multiple medicine knowledge graph.(2)Heterogeneity of medicine knowledge: medicine knowledge graphs contain not only the relationship between medical entities,but also rich attributes of various heterogeneous entities.To use the heterogeneous attributes in medicine evidence abstract,this thesis proposes a graph convolution attributed network based on dual attention mechanism.The main research contents of this thesis are as follows:(1)A cross-graph biased random walk method is proposed.For the complementarity of different knowledge bases,this thesis design the calculation methods of single graph node walk and cross-graph node walk probability,providing flexible hyperparameters and configurable scoring functions for controlling walk bias strategy.What’s more,this method try to incorporate node-node link information.(2)A graph convolution attributed network based on dual attention mechanism is proposed which include attribute-level attention and node-level hierarchical attention.For the underutilization of attribute features,attribute-level attention calculates the influence weight of different items.What’s more,node-level hierarchical attention calculates the influence weight of neighbors(direct and indirect neighbors).(3)Design and implement a medicine literature abstract extraction system Aceso2.0.The system integrates two models avbove which is used to help users upload literature and extract from PICO-related abstracts,and visualize the results in a structured and organized form.The results show that the method proposed can perform representation learning across medical knowledge graphs,and make full use of existing medicine knowledge and heterogeneous attributes,effectively improving the quality of medicine literature abstracts.