chunked/nltk_chunking/_attention/chunk_13.txt

This is shown in Figure 2d of the paper, see below for a sample attention mask:

Using those attention matrices with less parameters then allows the model to have inputs having a bigger sequence
length.