Research On Deep Learning Based Knowledge Graph Representation Learning Algorithm

The network has become a resource highland for information release and sharing with the popularization and development of Internet technology.Increasingly people use the network to release or collect information for various activities.How to effectively organize and utilize these data has become an important research position in the information age.Knowledge graph can effectively organize the knowledge in the human world into the form of data that can be processed by computers.The research of this paper aims to explore how to automatically construct effective knowledge graphs from massive text corpus data,and learn knowledge from knowledge graphs.The anchor node-based graph sampling method aims to solve the computational cost problem of massive nodes in large-scale knowledge graphs.The Conv Piece algorithm expands the node representation on the base of the anchor node-based graph sampling and optimizes the processing of context information.Finally,utilizing the method based on pre-trained language models to automate construct a knowledge graph and use the Conv Piece algorithm to complete knowledge graph representation learning.The main research contents of this paper are as follows:(1)This thesis proposes a graph sampling method based on anchors.In this method,the nodes on the graph are represented as the set of anchor nodes and neighborhood nodes.The main idea is that anchor nodes can represent the general information of the graph and neighborhood nodes can represent the unique information of the target nodes.The encoder is 8-heads single layer transformer encoder module for the sample-encoder model architecture.And proposes four different strategies for the construction of anchor nodes table.The influence of the four strategies on the accuracy of the node classification experiment was verified.The influence of anchor information and neighborhood information on the model performance was also analyzed in the ablation experiment.(2)This thesis proposes a deep-learning based knowledge graph representation learning method Conv Piece.In view of the problems of insufficient sampling granularity and insufficient strength in processing relational information in many current knowledge graph representation learning methods,this method proposes richer sampling content,uses two-dimensional convolution as an operator to process context information,and finally passes Transformer encoder and average pooling aggregated subgraph features form node embedding representations.Conv Piece is an encoder-decoder architecture.Parameters of various embeddings,encoder and decoder are trained end-to-end simultaneously.Link prediction experiments on two large-scale knowledge graph datasets,FB15k-237 and WN18 RR,showed that Conv Piece maintained 88% and 92% performance with only one-tenth the number of parameters of the larger model and outperformed the reference model by 9% and 2%,respectively.(3)In this thesis,an automatic construction method of knowledge graph based on pre-training large model is used to construct a knowledge graph for the American political domain from a large number of text corpus data in unsupervised mode.This thesis utilizes Conv Piece algorithm mining the knowledge contain in the constructed knowledge graph and proposed a relation reasoning model.The experiment results of visual knowledge reasoning show that the knowledge graph constructed in this thesis is of practical significance,and further prove that Conv Piece has the ability to explore unknown relationships between entities.