I think this is only on fake-scenario. In a real DMP, the user would be "loves cola","loves coca cola","loves fizzy drinks","drinks sugared drinks", and about 30 other related segments, and one of those would definitely sneak by.
Humans create audience segments, and it's in the interest of dataset producers to have each user be a part of the maximum number of members they can be to maximize monetisation.