chunked/nltk_chunking/model_doc_bert/chunk_3.txt

Unlike recent language representation models, BERT is designed to pre-train deep bidirectional
representations from unlabeled text by jointly conditioning on both left and right context in all layers.