Zero-shot Visual Question Answering with Language Model Feedback