| In the era of big data,knowledge graph contains a large number of factual entities and various relationships between them,which is a very important information base with wide application prospects in many fields such as recommendation systems,knowledge Q&A,and natural language understanding.However,existing knowledge graphs generally have incomplete data and a large number of missing relationships.The purpose of the link prediction task is to predict the missing relationships among entities in the knowledge graph,while the link prediction task is built on the basis of knowledge graph representation learning.The current translation-based and convolutional neural network-based knowledge graph representation methods mainly focus on each independent triad information,while the graph convolutional neural network-based methods can capture the deep semantic and structural information in the knowledge graph,but the existing knowledge graph sparsity is different,so in this paper,based on the existing research for the two types of knowledge graphs with different sparsity on the link prediction task Therefore,in this paper,we propose improved methods for two types of knowledge graphs with different sparsity based on existing research.(1)For large multi-relational knowledge graphs,since the relational information is rich enough,it is possible to characterize the entities by relations without considering the specific information of the entities themselves,and this paper proposes a codec structure model that fully considers the relational information.information identifies the relative positions between entities,reduces the time cost of the link prediction task by limiting the number of hops of the path,combines the relationship types and relationship paths in the coding part,and then uses a convolutional neural network-based decoder as a scoring function to evaluate the accuracy of the triples.(2)For sparse knowledge graphs,considering that the relationship information is sparse and it is impossible to portray entity features by relationship types only,while modeling only relationships will ignore the importance of different entities,so this paper proposes a joint neighborhood embedding model based on attention mechanism.The model embeds the entity-relationship neighborhoods around the target node separately first,assigns different weights to each neighborhood joint embedding using the attention mechanism,and finally aggregates the joint embeddings to obtain the target node embedding,and then uses several different decoders as scoring functions to compare the experimental performance.A large number of comparison experiments show that both model frameworks proposed in this paper have better performance compared with previous link prediction models on knowledge graphs with different sparsity. |
