Beyond VQA: Generating Multi-word Answers and Rationales to Visual Questions