Community Regularization of Visually-Grounded Dialog