I'd argue that the janky "queen = king – man + woman" thing is definitely related. It's not doing language generation, but you're trying to query the latent factors the model has used to minimise the surface objective
Word embeddings and latent factors in language models
By
–