Language models have the nice feature of being practically self-evaluating. We have introduced language models as a way of improving machine translation. But they are a small part of MT, and evaluating MT is difficult. (The best way is using them to translate the same set of sentences and simply asking a person which did the better job.)
