Comment by 100721

Does anyone know why they are using language models instead of a more purpose-built statistical model? My intuition is that a language model would either be overfit, or its training data would have a lot of noise unrelated to the application and significantly drive up costs.