Research And Implementation Of Completion Algorithm For Knowledge Graph

Knowledge graph completion refers to acquiring the graph features and mining the implied relationships based on the existing knowledge to construct a complete knowledge graph.To achieve automated knowledge graph completion,many methods have been proposed in recent years.Among them,neural network-based completion methods have received much attention due to their excellent nonlinear modeling ability and have performed well in extracting graph structure features and learning the representation of graph nodes.However,when these methods perform representation learning of the nodes of directed graphs,the nodes lack information about the directed relationships,and the information utilized is limited to complement the potential relationships among the nodes in the directed graph.At the same time,nodes contain information about neighboring nodes and relations at different distances,and the missing relations in the inference graph are also affected by the semantic information of neighbors at different distances,and it is not reasonable to use the traditional methods to obtain the semantic information of short distances to complete the graph.To address the above problems,this thesis improves the information aggregation method of the graph convolutional model and proposes an encoder-decoder model with the following main works:(1)In the directed graph,for the problem that the nodes lack the embedding information of the relations themselves when the traditional graph neural network learns the representation of the nodes,the 2D convolution operator is used to obtain the interaction information of the neighborhood nodes and the relational embedding,and the embedding information of the relations is fused into the nodes by this parameterized combination.For the directionality of relations in directed graphs,the relations are modeled separately according to different directions,and the representation learning of relations is realized to finally form an encoder model.(2)To obtain semantic information at different distances,convolutional kernels of multiple shapes are designed to extract richer semantic information.Also,in order to overcome the representation limitation of the shallow network model,deep convolutional neural networks are designed to extract deep semantic features and improve the completion effect.Finally,a decoder is formed,and the encoder forms an encoderdecoder structure for link prediction experiments.(3)Based on the improved model algorithm in this thesis,an encyclopedic knowledge Q&A system based on knowledge graph is designed and implemented,which can provide functions such as registration and login,knowledge Q&A,prediction and reasoning,and graph display,and the reliability and practicality of the model are confirmed in actual business situations.