Dynamic Graph Representation Learning for Video Dialog via Multi-Modal Shuffled Transformers.