The assumptions we make, the more performance goes down in tasks like machine translation, and that includes the assumption that parts of speech matter to machines.