Gravar-mail: Vision–Language–Knowledge Co-Embedding for Visual Commonsense Reasoning