The scientific case for embeddings is operational rather than metaphysical: they provide parameterised representations that make structured prediction computationally tractable. TransE, TransH, TransR, DistMult, ComplEx, ConvE, and TuckER encode different assumptions about relation geometry and interaction. No single architecture guarantees that Euclidean proximity equals human semantic similarity. Performance therefore depends on benchmark construction, negative sampling, calibration, and whether the graph’s relational patterns match the model’s inductive bias.