VideoBERT: A Joint Model for Video and Language Representation Learning