Research On Knowledge Graph Completion Model Based On Capsule Neural Network

Knowledge graph is a semantic network composed of many facts.These facts can usually be expressed in the form of triples(head entity,relationship,and tail entity),which are widely used in search engines,and intelligent assistants.Although the current knowledge graph contains a large amount of information,there is still a large amount of hidden information that has not been discovered.The purpose of knowledge graph completion is to solve this problem,explore the hidden knowledge in the knowledge graph,improve the completeness of the knowledge graph,and thereby improve the accuracy and efficiency of the knowledge graph.There are many ways to complete the knowledge graph,most of which are representation learning methods.In recent years,due to the rapid development of deep learning,many researchers use convolutional neural networks(CNN)to learn representations of knowledge.However,the knowledge graph completion model based on CNN has two defects: feature independence,the use of CNN can not capture the associations between the features in the knowledge(entity and relationship)vector;the pooling layer information Loss,the convolutional neural network pooling process will lose the characteristic information of the knowledge vector.In addition,in many current knowledge representation learning models,the number of feature interactions between entity vectors and relationship vectors is insufficient,which weakens the semantic relationship between entities and relationships in triples.Aiming at the two problems of CNN in the knowledge map completion model,this thesis proposes a knowledge map completion model based on the capsule neural network,namely the Caps-KGC model.The Caps-KGC model uses capsule neurons to encapsulate the features of entity vectors and relation vectors,so the features in the same capsule can be correlated with each other.At the same time,the model abandons the pooling layer operation and uses a dynamic routing algorithm to transfer the information between the capsule layers,so that all the feature information in the knowledge vector can be transferred to the next layer.Aiming at the insufficient number of feature interactions between entity vectors and relation vectors in the representation learning model,this thesis designs a knowledge graph completion model based on an improved capsule neural network based on the Caps-KGC model,namely the HCaps-KGC model.The HCaps-KGC model introduces the idea of a super network,and uses relation vectors to construct a convolution kernel to extract features of entity vectors.In addition,the HCaps-KGC model also proposes a method to enhance entity features.At the end of this thesis,experiments are carried out on the Caps-KGC and HCaps-KGC models.The experimental results show that the Caps-KGC model is superior to the existing models in most evaluation indicators.The improved HCaps-KGC model also has further improvements in various evaluation indicators.