chunked/nltk_chunking/model_doc_convbert/chunk_2.txt

However, BERT heavily relies on the global self-attention block and thus suffers
large memory footprint and computation cost.