TransE represents a relation as a translation between the vectors of two entities.

For a factual triple (h, r, t), TransE seeks a configuration in which h + r is approximately equal to t. Training compares observed triples with corrupted alternatives, adjusting vectors through an objective that makes valid configurations score better than invalid ones. This elegant translational geometry is efficient, but the retrieved literature notes limitations for one-to-many, many-to-one, and many-to-many relations, motivating models such as TransH and TransR.