Dan discusses extended semantic analysis using Wikipedia titles for text classification without training. He introduces response-driven learning as a way to improve meaning representation with minimal supervision, akin to reinforcement learning but with a single round approach.
