The first study I linked indicates that obfuscation is a little more effective than imitation, and both beat naive machine-translation output.
What if I attempt to transform a machine translation so that it imitates the style of someone well known?
I’d rather place the original work, machine translation, and a style-guide side-by-side and transform the original work to match the style-guide while also re-phrasing anything that tripped up the machine translation. That should cover both bases.
As someone who has not looked at the state of the art for stylometric fingerprinting since early in our eternal September, I wonder:
Naiively, that would seem like a strong approach. The translation would remove signal, and the imitation would add noise.
The first study I linked indicates that obfuscation is a little more effective than imitation, and both beat naive machine-translation output.
I’d rather place the original work, machine translation, and a style-guide side-by-side and transform the original work to match the style-guide while also re-phrasing anything that tripped up the machine translation. That should cover both bases.