| 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. |