Research On Knowledge Graph Embedding Method

The knowledge graph is a structured representation of real world facts.As the name suggests,knowledge graphs model the connections of everything in the real world in the form of graphs,but they usually only contain a small part of all possible facts.Due to the incompleteness of the knowledge graph,many applications based on the knowledge graph are restricted,such as: intelligent question answering systems,auxiliary medical care,recommendation systems,etc.Furthermore,the task of completing the knowledge graph is generated.In order to automatically complete the knowledge graph,the elements in the knowledge graph need to be represented by continuous vectors,so the knowledge graph embedding appears.Therefore,how to embed the knowledge graph and making the model complete the knowledge graph more accurately and efficiently is particularly important.The facts in the complex knowledge graph can be well represented by tensors,and tensor decomposition has strong expressive ability and explicable.Therefore,many mod-els are based on tensor decomposition for knowledge graph embedding.In addition,the embedding model based on convolutional neural network is also a research hot.Although these models solve some problems in the knowledge graph embedding and come up with new ideas,they fail to fully learn the semantic information.There are the following two parts in this thesis:First,the classical knowledge graph embedding model,whether it is an efficient model based on tensor decomposition,or a simple model based on translation,ignores the component semantic information of entity.To solve the problem,this thesis proposes a model based on tensor decomposition that can learn paired knowledge.The model em-beds the component semantic information of the entity in the space of the relationship,which enables the method capture the component semantic information of the entity in re-lationships embedding space,and retain the simplicity and efficiency.And then develop the link prediction experiments on the data sets FB15k-237 and WN18 RR.Each evalua-tion index is very good,which verifies the performance of the model including component semantics.Second,in view of the defect that the classic models based on convolutional neural network can not effectively learn the semantic information of the fusion of entities and relationships,the thesis proposes the convolutional neural network based on tensor prod-uct.Firstly,the interaction semantics of entities and relationships are captured through tensor product,and at the same time,the one-dimensional vector is shaped to a matrix.Then,immediately after the tensor product,a convolutional neural network follows.Fi-nally,a link prediction comparison experiment is carried out on the benchmark data sets FB15k-237 and WN18 RR.The experimental result shows that the new neural network,capturing interaction semantics between entites and relationships based on tensor prod-ucts,can complete the completion task more accurately,which proves the effectiveness of tensor product in semantic fusion.