Learning Entity and Relation Embeddings for Knowledge Graph Completion

The paper targets knowledge graph completion, which performs link prediction between entities, using the approach of knowledge graph embeddings. Earlier translation-based models such as TransE and TransH represent a relation as a translation from the head entity to the tail entity, but they embed both entities and relations within a single semantic space. The authors argue that an entity has multiple aspects and that different relations focus on different aspects, so one shared space is insufficient. TransR therefore builds entity and relation embeddings in separate entity and relation spaces, projecting entities into the corresponding relation space before building translations.

Evaluated on three tasks, link prediction, triple classification, and relational fact extraction, TransR delivers significant and consistent improvements over state-of-the-art baselines including TransE and TransH. By modeling entities and relations in distinct spaces, the method captures relation-specific aspects of entities more faithfully, and it became an influential baseline for subsequent knowledge graph embedding research.