114 - Behavioral Testing of NLP Models, with Marco Tulio Ribeiro
NLP Highlights
00:00
Using the Paraphrases in a Logic Model?
Our n o p models are not at all like soft ware, in that we can have unit tests for soft ware because the execution of the functions is independent. But if i introduce some correlate that the model tries toe som some confounding factor that the model actually doesn't disentangle the way you might think it should, then all of a sudden the model fails on nication. So one change that you would make, wrote one paper which was just about one kind of perturbation, is to do paraphrases and say, the motto should be robust to paraphrases. And so had an example of adding very negative phrases to the end of a review. I hate this.
Transcript
Play full episode
