chunked/nltk_chunking/model_doc_swiftformer/chunk_6.txt

Although hybrid approaches have been proposed to combine the advantages of convolutions and self-attention for a better speed-accuracy trade-off, the expensive matrix multiplication operations in self-attention remain a bottleneck.