Embedding training learns vectors by contrasting true triples with corrupted triples.

A typical procedure initializes entity and relation vectors, samples batches of known triples, and creates negative examples by replacing a head or tail with another entity. An optimisation process then updates the embeddings according to a scoring function, increasing the plausibility of observed triples relative to corrupted ones. The resulting vectors can be used to rank candidate links, but their semantic interpretation depends on the model, training graph, loss function, and evaluation protocol.