We first examined the representation of the 20 languages in M-BERT by deriving language identity representations in 1000 labeled corpora. The high performance of the language identification model in distinguishing the ...