The genres come from the different models AcousticBrainz uses, https://acousticbrainz.org/datasets/accuracy gives some details about the different genres as defined per model. The API returns raw values here as used by the respective model.
I agree, it would definitely make sense to translate those in cases where it is displayed to the user. E.g. the AcousticBrainz Picard plugin should translate those values to something better readable. This would also give the chance to unify the genres where two models agree, so e.g. instead of “rock; roc” it would become just “rock”. Maybe even sort the resulting list by how many models found the same genre.
EDIT: For the Picard plugin I added an issue at https://github.com/metabrainz/picard-plugins/issues/228 . So if someone wants to pick this up please do