Research On Knowledge Graph Representation Learning Based On Attention Mechanism

In recent years,with the rapid development of artificial intelligence and Internet technology,it has become increasingly important to extract high-value information from massive data accurately and efficiently.As a semantic knowledge base that stores knowledge in a structured way,knowledge graph uses graph model to represent entities in the real world and the relationships between entities,which can help people to quickly conduct knowledge retrieval and knowledge reasoning.Aiming at the problem that the knowledge representation learning model based on graph neural network indirectly aggregates higher-order neighborhood information by means of stacking multi-layer network,which leads to the loss of higher-order neighborhood information and noise,A knowledge representation learning model named Knowledge Embedding Based Graph Attention Diffusion(KE-GAD)is proposed in this paper,which can directly aggregate higher-order neighborhood information in single-layer network.KE-GAD model improves the Graph Diffusion Convolution(GDC)model by adding attention mechanism to the GDC model,so that it can use attention mechanism to directly aggregate higher-order neighborhood entity information in single-layer network,and add context awareness mechanism to the GDC network.Learn entity embedding vectors that fuse higher-order and lower-order neighbor entity information.To solve the problem that GDC model has no relational learning mechanism,a relational learning mechanism is added to make it suitable for knowledge graph representation learning.In this paper,we conduct node classification and link prediction experiments on the KE-GAD model on multiple data sets.The experimental results show that the KE-GAD model can achieve better results than the model using only entity information.In order to solve the problem that the learning effect of knowledge representation is not ideal when the KE-GAD model is compared with the model that integrates other information,this paper proposes a multi-information fusion representation learning method based on attention mechanism(Multi-Info KE).Firstly,based on the attention mechanism,the method encodes the triplet’s translation feature information,the relational information and the triplet’s description information to obtain the entity feature vector that integrates the three kinds of information.Then,in order to aggregate higher-order neighbor information,the three entity feature vectors obtained after coding and the entity feature vectors input into the network were respectively input into KE-GAD model,and the higher-order neighborhood information was aggregated and the entity embedding vector and relational embedding vector were learned.Finally,the entity embedding vector and relational embedding vector output by KE-GAD model are fed into the knowledge representation learning decoder for information fusion to generate the final entity embedding vector and relational embedding vector and calculate the triple score.In this paper,we verify the information coding ability of multivariate information coding module on WN18 RR and FB15K-237 data sets,and the effect of fusion of multivariate information in knowledge representation learning.The experimental results show that the fusion of multiple information can effectively improve the learning effect of knowledge representation.