GPT-3 can interpret critiques in a positive light!

Hi All,

Just sharing our recent short paper (preprint: [2109.07576] "It doesn't look good for a date": Transforming Critiques into Preferences for Conversational Recommendation Systems) accepted at EMNLP.

Here, we investigate whether GPT-3 can interpret free-form critiques to restaurants (e.g., “It doesn’t look good for a date”) as positively stated preferences (e.g., “I prefer a more romantic place”) in order to retrieve better recommendations (e.g., “This place is perfect for a romantic dinner”). As described in the paper, our prompt was very simple: 10 conditioning examples + the new critique + “I prefer”.

We found three types of critique-to-preference inference that GPT-3 can perform:

  1. When the user implies a preference for a feature using the polar opposite (e.g.,
    “It looks too casual” implying “I prefer a fancier place”);
  1. When the user draws on common sense to express a preference (“It has a freaking band!” implying “I prefer a more quiet place”); and

  2. When the user implies a filter within a set of related features (e.g., “I don’t really like seafood” implying preference for alternatives in the meat category, or “I don’t eat meat” implying preference for vegetables).

Besides the ability to interpret non-trivial phrasings such as “How come they only serve that much?” as “I prefer larger portions.” In case you are interested in critique-to-preference inferences, many other successful examples can be found here under the tab “Characterization of Differences.”

Overall, we found that adding this simple interpretation given by GPT-3 improves the retrieval of appropriate customer reviews by 19-124% depending on the technique, which could translate into better recommendations and more natural interactions with conversational recommender systems.

I’m happy to discuss it further if you have any questions or feedback!
Many thanks to this community for the amazing work,