The Dialog Must Go On: Improving Visual Dialog via Generative Self-Training

Section A shows a detailed model architecture. • Section B presents further quantitative analysis. • Section C presents further qualitative analysis. • Section D presents implementation details. • Section E shows discussion, including limitations, future work, and ethical considerations.