I am playing with a large dataset of hotel reviews, which contains both positive and negative reviews (the reviews are labeled). I want to use this dataset to perform textual style transfer - given a positive review, output a negative review which address the same thing. For example, if the positive review mentioned how spacious the rooms are, I want the output to be a review that complains about the small and claustrophobic rooms.

However, I don't have positive review-negative review pairs for the training. I was thinking that maybe I could create those pairs myself, but I'm not sure what is the best way to do that. Simple heuristics like jaccard index and such didn't give the desired results.

  • 1
    $\begingroup$ This seems like a tricky data set to create automatically. Do you have a different goal you want to eventually achieve, once you have this data set? Put differently, why are you building such a data set? By the way, "style transfer" can very often be equated with "translation" in NLP. $\endgroup$ – Mathias Müller Jan 30 '20 at 19:43
  • $\begingroup$ Well, my goal is to show that supervised style transfer can work in practice in complex cases, so I could get funds for labeling a very challenging dataset :) $\endgroup$ – Nadav Borenstein Feb 3 '20 at 7:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.