Research And Application Of Knowledge Graph Completion Method Based On Graph Structure Relationship Reasoning

With the advent of the era of big data and the vigorous development of artificial intelligence technology,since the concept of knowledge graph was formally proposed in2012,research on knowledge graph technology has been continuously updated iteratively.The knowledge graph is everywhere,and smart cities,intelligent retrieval,question answering systems,etc.are closely related to our daily lives.However,in the development of knowledge graph,there have also been many difficult problems to solve,such as lack of integrity of knowledge graph,and low content of node information.This thesis aims to improve the existing knowledge graph inference technology against the above shortcomings.The analysis is conducted from two perspectives: multi-level local subgraph reasoning based on inductive relationships and global path reasoning based on self attention mechanism.Finally,the encoded information output from the two modules is organically combined.Finally,a mental health early warning system based on knowledge graph reasoning is designed,which applies knowledge graph reasoning to the field of mental health.The main work of this thesis is as follows:(1)Aiming at the problems of current graph neural networks,such as a single information transmission mechanism and less utilization of high-level semantic features of subgraph structures,a node edge bidirectional enhancement mechanism is proposed to update nodes and edges to obtain richer contextual semantic information.High level graph neural networks are used to model multi-level subgraphs,fully capturing high-level semantic features of subgraph elements of different levels,Be able to conduct inductive reasoning on entity independent relationship information in a subgraph.The effectiveness of node edge bidirectional enhancement and hierarchical subgraph construction was verified through experiments.(2)The current path reasoning model has problems such as long-term dependence and insufficient modeling ability of path sequence information.The Transformer encoder is used to model the extracted relational paths,and the dependency relationships between all entities are obtained through a self attention mechanism,allowing more logically related paths to have higher weights,improving the performance and accuracy of relational reasoning.The effectiveness of path relationship encoding,path fusion,and joint reasoning was verified through experiments.(3)Design and implement a mental health early warning system based on knowledge graph reasoning as an application of the algorithm proposed in this thesis.This system uses the field of mental health as an application background.Aiming at most people with varying degrees of mental health problems,especially depression and suicidal tendencies,and the high cost of offline mental health treatment,it builds a mental knowledge graph by capturing and processing cross modal data,and applies the reasoning algorithm model proposed in this thesis to the mental health knowledge graph,improving the early warning efficiency of mental health,It meets the practicality in real life.