I don't want to say "recipe for disaster", but as someone working on a similar app, it's definitely a challenge. It's one of the reasons I stayed away from a pu

I don't want to say "recipe for disaster", but as someone working on a similar app, it's definitely a challenge. It's one of the reasons I stayed away from a pure ML/language modelling approach. It's really easy for models to get things wrong - for example, elsewhere in this thread there's a Chinese sentence about arranging furniture but the word "arrange" isn't really right (it's more "arrange" in the sense of "arranging a delivery", rather than "arranging a number of objects", and sounds like it's been auto-translated from English). When I started I wanted to be sure that every single word/sentence was "correct" (i.e. a native speaker agreed that all the words/grammar were correct/natural in that context). LMs have improved since I started, but for me, that requirement still stands and ML approaches still need a heavy human touch IMO.