75 - Reinforcement / Imitation Learning in NLP, with Hal Daumé III
NLP Highlights
00:00
The Future of Semantic Parsing
The banded structured is a little harder because you only get to evaluate one, but you can still compute a lot. I think that semantic parsing and program synthesis are really awesome problems to think about in this space. There's a lot of room for both building better imitation or reinforcement algorithms and hopefully getting better at this problem. The re-slope idea is basically to try to automatically do this sort of credit assignment. This sounds a lot like inverse reinforcement learning.
Transcript
Play full episode
