In the lecture @jeremy says in you need to train your domain language model once; like it might take overnight or 2-3 days to train. Once it is done, we can crate different kinds of models and classifiers on top of that. Explain me what is happening, why and how?
Because this becomes similar to our ImageNet model we transfer learn with. (Also recommend not @‘ing Jeremy if it’s a general question). If the language we expect to show up in production shifts at all then perhaps we may consider retraining this language model to fit our adjusted corpus. But from there we train the classifier. The NLP course goes into much detail over this.