NLP - Extract support phrases for sentiment

Hi all,

I’m trying out an NLP competition on kaggle, and can’t figure out what next steps to try. The goal is to extract support phrases for sentiment. I tried hacking the text_classifier_learner to do multi-labeling for the output, but I couldn’t get it working and am starting to think there has to be a better way.

Input: I am having a good, fun time at home.
Output: good, fun
Sentiment: positive

Does anyone have recommendations on notebooks/posts/examples on how to do this task with fastai? I’m trying to figure out how deep into the fastai API I need to dig for this, and where to dig.

The competition is here.

Added 5-15-20

  • Noticed there was a text interpretation with intrinsic attention in fastai 1v3

Would any veterans know if it which is more worthwhile: porting 1v3 over, or try to use hugging faces for the task? I saw some folks are already experimenting with hugging faces with fastai2 (1v4).

Thank you,
Daniel Lam