The large-scale language model then identifies the grammatical role of each token and infers the subject, verb, object, etc. of the sentence.