Indeed, I struggle to even imagine what "use unusual language when describing well-known celebrities" even means! Maybe like using "musician" rather than "artist" or some other combination?

edit: Ok, I've read through the paper, and still have no idea. Apparently the responses to questions were compared as semantic vectors using cosine similarity in Google’s Universal Sentence Encoder space. Or something lol.

and interestingly they say they share their data, but after looking through the data I don't see what I'm looking for, which is closest-approximate words for each celebrity.

"hello fellow Taylor Swift fans"