[dekhn]

It didn't pick "something"- it chose scientific nomenclature as a basis, and synthesized new classes from that basis.

They're not nonsense words, they're words with high probability which are not seen in the dataset.