chunked/nltk_chunking/model_doc_convbert/chunk_4.txt

We therefore propose a novel span-based dynamic convolution to
replace these self-attention heads to directly model local dependencies.