It works on embeddings, which are concepts. So it would at least be manipulating a similarly weighted network of concepts, though much shallower. It would be an answer from his literal domain of discourse. It would be evocative rather than authoritative.