Sure, use character n-gram or word-part models that generate logos from input text that doesn’t have to be matched to a fixed vocabulary. When input text is weird, it should never fall back to trying to make the user self-select into a category and the logos should not try to be generated from coarse category information at all.