(Not OP) I can think of a convoluted and expensive pair-wise comparison method, but I hope there's also a way to figure this out during the application of principal component analysis in a way I don't understand.

Edit: I'm thinking it can't be done without experimentation on the embedding model.

Edit2: Ah, even that might not yield results, because as the basis is derived interstitially through computation, there's no guarantee the features of the final coordinate system will have any accessible relationship to those of the initial basis.