chunked/nltk_chunking/model_doc_encodec/chunk_5.txt

Finally, we study how lightweight Transformer models can be used to further compress the obtained representation by up to 40%, while staying faster than real time.