Explaining VQA predictions using visual grounding and a knowledge base