Knowledge Graph Completion Based on the Joint Embedding of Entities

Abstract

As the last step in the process of knowledge graph construction, knowledge graph completion predicts and inferences new factual triples according to the existing knowledge, which is of great significance to the downstream application of the knowledge graph. The knowledge graph completion technology learns representations of entities and relations in the knowledge graph and reconstructs triples factual relations in the representation space to measure their effectiveness, so as to predict and supplement entities and relations. At present, Graph Convolutional Network (GCN) is widely used in entity modeling, which combines the graph structure information to learn the vector representation of entities. However, most algorithms only focus on static knowledge graph completion, ignoring the important role of time in the knowledge graph, and the modeling of entities is not comprehensive enough. In this paper, a knowledge graph completion based on the joint embedding of entities is proposed, which combines structural feature embedding of the graph and dynamic embedding of entities to carry out multi-level learning of entities to improve the effect of entity prediction. Structural feature embedding of the graph takes advantage of the powerful feature aggregation capability of GCN to integrate the knowledge graph as the topological structure and node features of graph data into the representation learning of entities. Dynamic embedding of entities takes advantage of the characteristics that entities will behave in different states at different times and divides the entity features into two parts for learning. One part of the features changes dynamically over time and the other part retains its static properties. In this paper, the probability model ConvTransE is used to construct the model of relations and triples, ConvTransE uses the convolutional neural network to extract the semantic information of the triples and predict the missing elements. We have conducted experiments on public datasets and achieved better results on Hit@K and MRR.