Choosing a domain zone by image?

Question

Machine learning

It is necessary to select a domain zone from newGTLD according to the image. For example - a girl in a swimsuit - .SEXY domain, a person with a glass of wine in the photo - .WINE, a person in a business suit - .BUSINESS. At the moment, the prototype is based on the Google Vision Api, the resulting set of image features is compared for similarity with a manually compiled dictionary of synonyms for each domain zone using word2vec. The results are not very subjective. So, for a girl in a bathing suit, the obtained features are: "human hair color", "girl", "leg", "interaction", "mouth", "black hair", "organ", "finger", "long hair", " thigh" are semantically very far from being synonymous with the word sexy.
What other approaches can be tried to solve such a problem? About 500 domain zones.

Answer 1

I think because the answer is very subjective. Why does a girl in a swimsuit have to be "SEXY" and not "swimsuit" or "beach"?
Approaches to the solution - we take a large body of pictures, manually place them with those tags that we consider correct, build a model and then use it.