Enhancing knowledge graph embedding with type-constraint features

Knowledge graph (KG) embedding represents entities and relations with latent vectors, which has been widely adopted in relation extraction and KG completion. Among existing works, translation-based models treat each relation as the translation from head entitiy to tail entitiy and have attracted much attention. However, these models only utilize fact triples but ignore prior knowledge on relational type-constraints. This paper presents a generic framework to enhance knowledge graph embedding with type-constraint features (ETF). In ETF, the embedding of entity is comprised of two parts-entity-specific embedding and constraint-specific embedding. The former expresses translation features of entities, and the latter represents semantic constraints influence by linked relations. Besides, the adaptive margin-based loss is designed to learn embeddings, which effectively separates the negative and positive triples. Finally, the results on four public datasets demonstrate that ETF makes significant improvements over the baselines.