Knowledge Graph Construction And Reasoning Based On Graph Neural Network

This thesis has been optimized and improved from three parts: knowledge representation in the process of constructing knowledge graph,knowledge extraction problem and reasoning task.The topic first studied the knowledge representation and visualization of learner comments in the construction of educational knowledge graphs.By selecting learners' comment on education big data as the knowledge source,the knowledge representation in the multi-dimensional semantic space is explored.Second,the study expanded the definition of learner comments,and proposed to define the entities and links in its semantic model from the time dimension,emotional dimension,and social dimension.By introducing a series of algorithms and formulating a data extraction mechanism,the study realized the construction of the knowledge graph of the learner's comments.Furthermore,the study explored the above three kinds of semantics by fusing sentiment analysis algorithm,social semantic analysis and time series analysis algorithm,and realized the visualization of knowledge graph that is convenient for learning and analysis.The second part of the topic mainly involves humanity and knowledge graph construction methods and construction strategies.It is different from the previous knowledge construction methods.In the construction of ancient poetry knowledge graph,knowledge that is more difficult for human cognition is extracted in the process It is often unable to extract accurately,and has strong human intelligence dependence and data labeling requirements.By designing a reasonable group intelligence and knowledge graph task allocation mechanism,the efficiency of knowledge graph construction for this type of labeling task can be improved.This part mainly optimizes the classification of knowledge graph tasks,the distribution process and the question scoring mechanism from the three levels of crowdsourced knowledge graph distribution,reverse verification code,and group intelligence representation algorithm.The optimized group wisdom construction can reflect more human cognition and past labeling situations when constructing the knowledge graph,and extract higher-quality knowledge triples.The third part of the subject is mainly the reasoning module of the knowledge graph.This module mainly talks about link prediction tasks.Different from the previous reasoning models in knowledge graphs,such as distance-based models,semantic models,and neural network models,the graph neural network method is used in the knowledge reasoning construction process of this research.Since the graph neural network method can better represent the graph data characteristics of the knowledge graph,we chose the DD-GCN optimization model that adopted R-GCN and developed the fusion degree information difference measurement as the baseline model of the task.Experiments show that DD-GCN can improve the inference task effect of R-GCN's graph neural network in the related tasks of entity classification and link prediction.In general,the subject explores and improves the multi-dimensional knowledge representation of the knowledge graph,the construction of group wisdom,and the graph neural network reasoning from multiple angles,which has certain reference significance for the development of knowledge graph construction and reasoning.