1
$\begingroup$

Sentiment analysis, as we know, measures "Cake sucks" as say -0.4, and "Cake is great" as 0.7. What I'm looking for is something a bit different like so:

  1. Given input text data written by 1 person (say a blog)
  2. Predict how they (the person who wrote the text) might react to a certain piece of text

What might something like this look like?

  • Let's suppose that Person A with a blog has written in his blog posts thousands of times about how much cake is the best thing to happen to humanity.
  • The system should probably infer that if that person read something like "Cake is the WORST food ever", they would react negatively to it, if say, they also believe that there is such a thing as 'objective taste' somehow (aesthetic absolutism).
  • Or if Person A has made anti-racist statements, that racist statements would be strongly negative.
  • If Person A reads the statement "I hate lawyers" and in their blog they have written about how they don't care either way about law, it should probably be 0.
  • Finally, if Person A reads the statement "iPhones are better than Android" and there is zero data about either iPhones or Androids, or even related data about Apple or Google, then it should probably be 0, with an additional "confidence" metric at 0 (since there is no data, this confidence metric will let us know whether there is any data to support the measurement or not).

This model would need to be able to somehow inductively 'infer' a value system of some kind, and assign intensities of probable reactions based on the frequency of an expressed view, as well as pick up on nuances (such as philosophical assumptions, (for example in the above cake example: aesthetic absolutism) etc.) that may inform that measurement.

In other words, I'd like to create a model (or find a pre-trained model to fine-tune), that would be able to, given text data from that 1 person, predict their sentiment in response to a new piece of text.

Would love any help whatsoever regarding:

  1. What types of pre-trained models I should look at
  2. Any ideas of any kind whatsoever you might have on how to achieve this
  3. What sorts of architectures/resources/concepts may be relevant to look at
$\endgroup$

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.