chunked/nltk_chunking/model_doc_albert/chunk_5.txt

To address these problems, we present two parameter-reduction
techniques to lower memory consumption and increase the training speed of BERT.