Textbook Question Answering with Multi-modal Context Graph Understanding and Self-supervised Open-set Comprehension