Reasoning Visual Dialogs with Structural and Partial Observations