marcotcr / checklist

Beyond Accuracy: Behavioral Testing of NLP models with CheckList
MIT License
2.01k stars 204 forks source link

Text Generation model for negation #110

Closed emc5ud closed 2 years ago

emc5ud commented 2 years ago

Hello, I was considering working on a model that generates negated versions of simple (for now) input sentences. I was wondering if you thought that this could be a good feature for the Perturb class, or if you would prefer to keep the negation perturbations rules-based for now. Maybe there could be options to use one or the other?

If you have ideas for some other more low hanging fruit I could help with, please let me know! I'm interested in contributing.

marcotcr commented 2 years ago

If you have a model that is not too big and is high-precision, I would be happy to replace the rules with it. You can take a look at this other repo for a counterfactual perturbation model, but it would be some work to make sure it's high precision. I don't think this is low hanging fruit, but I could be wrong :)