How do we know whether a language model is good? There are two basic approaches: Task-based evaluation (also known as extrinsic evaluation), where we use the language model as part of some other task, like automatic speech recognition, or spelling correcktion, or an OCR system that tries to covert a professor's messy handwriting into text. Intrinsic evaluation: Intrinsic evaluation tries to directly evalute the goodness of the language model by seeing how well the probability distributions that it estimates are able to explain some previously unseen test set