123 - Robust NLP, with Robin Jia
NLP Highlights
00:00
How to Defend Against Adversarial Typos
We were able to do this specifically in the case of trying to defend against adversarial typos. So previous work looked at things like using a typo corrector, which like makes, I think intuitively makes makes a ton of sense. And so if you have some sort of adversary that's trying to insert a couple of typos to fool the system, they can fool the typo corrector into doing something weird and then that'll cause the model to do something weird. We use this to get kind of much better robustness to adversarial chosen typos than from previous work. The key thing is just that we want these encodings to be these like discrete objects where basically like if
Transcript
Play full episode
