Overcoming Language Priors for Visual Question Answering Based on Knowledge Distillation