chunked/nltk_chunking/model_doc_funnel/chunk_4.txt

With this intuition, we propose Funnel-Transformer which
gradually compresses the sequence of hidden states to a shorter one and hence reduces the computation cost.