Yes, what I'm saying is that gzip does not perform as well when it's not overlapping tokens exact. Gzip does not support a "semantic" mode, hence it won't and

Yes, what I'm saying is that gzip does not perform as well when it's not overlapping tokens exact. Gzip does not support a "semantic" mode, hence it won't and does not (according to the papers metric) perform as well. Deep learning can capture these semantic similarities.