I am currently examining bias in NLP datasets such as CommonsenseQA. I am looking for a reference where researchers possibly claim at which point, i.e., how biased a dataset must be, in order for an ML model to pick it up. Is anyone aware of work going in this direction? I cannot find anything.

  • $\begingroup$ It starts at one data point. (What does it even mean to say “how biased” the data is? There’s a good quote from E.B. White, oddly enough, about how all data are biased.) $\endgroup$ Aug 11, 2021 at 13:47
  • $\begingroup$ What does data bias mean here? Maybe read en.wikipedia.org/wiki/Bias%E2%80%93variance_tradeoff $\endgroup$
    – Dirk N
    Aug 11, 2021 at 15:26


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.