Knowledge Graph Embedding by Translating on Hyperplanes

The paper tackles embedding a large knowledge graph of entities and relations into a continuous vector space. It observes that TransE, though efficient and highly accurate, struggles with important relational mapping properties—reflexive relations and one-to-many, many-to-one, and many-to-many relations—whereas richer models capture these but lose efficiency. To balance capacity and efficiency, the authors propose TransH, which models each relation as a hyperplane together with a translation operation on that hyperplane, preserving the mapping properties at almost the same complexity as TransE. They also introduce a simple trick that exploits a relation's mapping type to construct negative examples that reduce false-negative labeling during training.

Across benchmark datasets including WordNet and Freebase, and on tasks of link prediction, triplet classification, and fact extraction, TransH delivers significant improvements in predictive accuracy over TransE while retaining comparable ability to scale up. This showed that carefully modeling relation-specific projections can resolve TransE's limitations on complex relations without giving up its efficiency.