Knowledge Graph Embedding Based on Joint Entity Types and Relational Compositions for Knowledge Graph Entity Completion

Knowledge graph is a handy tool to show how things are related to each other. To facilitate computation and reasoning on the knowledge graph, embedding techniques are needed to map the knowledge to a lower dimensional space. However, the importance of entity types and information about the different locations of entities is often underestimated in knowledge graph embedding techniques. To address the shortcomings of current embedding methods, we propose an embedding framework for knowledge graph entity completion based on entity types and relational compositions, noted as JETRC. Our framework integrates type information through an entity type transition matrix and the use of relational composition representations to learn the latent information of entities at different locations in triple, thus better capturing the semantic relationships and logical dependencies among entities. An attention mechanism-based encoder is used to update embeddings of entities and relations. Experimental results indicate that JETRC significantly improves accuracy in the task of knowledge graph entity completion. It outperforms the comparison model in MRR metric, Hits@1 metric and Hits@10 metric.