Generating plausible but incorrect answers/translations?


I know generating plausible incorrect answer has been an issue. I want to generate 1. useful English phrases, 2. translations into my language, and 3. slightly incorrect translations.

It should be clear which translation is correct only after reading the original phrase.

The model usually generates sentences which are obviously wrong in terms of logic OR translations with incorrect grammar OR sentences which are different, but also correct translations.

What’s your approach to generating incorrect, but believable anwers in Q&A sets?