As far as I know, DeepL is allowed to (and will prioritize) naturally-sounding sentences so it feeds in the extra context it received during ML model training. There are some cases where the answer is extremely specific (f.e. a name of a particular institution) which shows that the data set was oddly unbalanced.
I haven’t think of this much before, but there is a possibility of using DeepL to translate to Japanese (with a grain of salt, nonetheless).
I can’t yet find a fitting vocabulary for vocabulary hunting, though.
You and me both Google. I’d say someone needs to whip Google into shape but from the sounds of it… it might enjoy it.