Further analysis demonstrates that VisualBERT can ground elements of language to image regions without any
explicit supervision and is even sensitive to syntactic relationships, tracking, for example, associations between
verbs and image regions corresponding to their arguments.