chunked/nltk_chunking/model_doc_cvt/chunk_3.txt

This is accomplished through 
two primary modifications: a hierarchy of Transformers containing a new convolutional token embedding, and a convolutional Transformer 
block leveraging a convolutional projection.