Overcoming Language Priors in Visual Question Answering with Adversarial Regularization