Research And Application Of Key Technologies For Knowledge Graph Completion Based On Data Enhancement

As a carrier for structured representation of knowledge and storage of knowledge,the knowledge graph gradually provides data support and reasoning decisions for various search engines,question and answer systems,etc.It is a core key technology for artificial intelligence.Knowledge graph completion technology is one of the hot core problems in knowledge graph research.Knowledge graph completion refers to the process of adding missing entities and relationships to the knowledge graph or updating new knowledge to the graph.The current knowledge graph completion technology mainly relies on encoding entity structures for completion,but the research on completion technology still faces many challenges.The main challenges include:(1)Incomplete information exploitation.Knowledge graph completion methods only make use of entity structures for completion,and do not fully consider the problems of too few data samples,long-tail distribution of relationships and structural heterogeneity in the knowledge graph;(2)The completion model is not robust.The unstable quality of the knowledge graph data leads to the poor robustness of the completion method.These problems seriously restrict the application effect of the knowledge graph completion technology.To address the two major scientific problems of incomplete information utilization such as structural heterogeneity and sparse samples as well as unstable data quality and poor model robustness in the knowledge graph completion technology,completion methods based on data enhancement are proposed in this paper,specifically including:(1)For the structural heterogeneity and sample sparsity problems,a sample and feature enhancement method is proposed.The method first uses generative adversarial network to learn the parameters to generate structural and content features of entities,and increases the number of samples as well as the feature richness of entities during the training process by training mature generators;the method uses the content features of entities to compensate for the defects of structural heterogeneity and balances the expression weights between them through an adaptive mechanism to solve the problem of structural sparsity in knowledge graphs.In order to verify the effectiveness of the data-enhanced knowledge graph completion method,the paper conducts completion experiments on five real knowledge graph datasets,such as Wiki and NELL.The experimental results show that the proposed method is significantly more effective than other methods,which validates the effectiveness of the data-enhanced method.(2)To address the problems of unstable data quality and poor robustness of the completion methods,the paper proposes a method based on convolutional coding and self-attentive fusion.The paper analyses the robustness of each completion method under knowledge noise and data perturbation scenarios,and proposes convolutional coding of entity content features and fusion using a self-attention mechanism to address the problem of poor data quality.To verify the robustness of the completion model in the presence of data noise scenarios,the thesis conducts completion experiments on three real datasets such as NELL with artificially designed noise,and the results show that the methods in the thesis have better robustness compared to other methods.The paper proposes an improvement method based on data enhancement to address the problem of poor data quality.The experimental results show that the data enhancement-based method can solve the problem of data noise in the domain of knowledge graphs.(3)Integrating the research results of the article and applying them to build a medical knowledge graph Q&A application.Based on the chinese medical knowledge graph,the paper crawls relevant medical information,uses the model in the paper to perform the completion and verifies the model robustness.By comparing and analysing the effect of the completion,the paper finds that the data enhancement-based approach can solve the problems of incomplete and untimely updates of the medical knowledge graph.The paper applies the completion method to the implementation of the medical question and answer system.In the process of building the medical question and answer system,the thesis applies the method of completion to complete and update the knowledge.By comparing and analysing the feedback before and after the Q&A system,the thesis verifies the practicality of the data-based augmentation method in real-life scenarios.