chunked/nltk_chunking/model_doc_xlnet/chunk_3.txt

However, relying on
corrupting the input with masks, BERT neglects dependency between the masked positions and suffers from a
pretrain-finetune discrepancy.