Lecture
This lecture covers the challenges in natural language generation, focusing on decoding algorithms such as greedy decoding, beam search, and sampling methods like top-k and top-p. The instructor, Antoine Bosselut, explains the training challenges including exposure bias and reinforcement learning. Various solutions to improve decoding, such as re-balancing distributions and re-ranking sequences, are discussed. The lecture emphasizes the importance of diverse text generation and the use of reward functions in reinforcement learning. Different reward estimation metrics like BLEU and ROUGE are explored, along with behaviors tied to rewards like cross-modality consistency and utterance politeness.