ISAAQ -- Mastering Textbook Questions with Pre-trained Transformers and Bottom-Up and Top-Down Attention

Textbook Question Answering is a complex task in the intersection of Machine\nComprehension and Visual Question Answering that requires reasoning with\nmultimodal information from text and diagrams. For the first time, this paper\ntaps on the potential of transformer language models and bottom-up and top-down\nattention to tackle the language and visual understanding challenges this task\nentails. Rather than training a language-visual transformer from scratch we\nrely on pre-trained transformers, fine-tuning and ensembling. We add bottom-up\nand top-down attention to identify regions of interest corresponding to diagram\nconstituents and their relationships, improving the selection of relevant\nvisual information for each question and answer options. Our system ISAAQ\nreports unprecedented success in all TQA question types, with accuracies of\n81.36%, 71.11% and 55.12% on true/false, text-only and diagram multiple choice\nquestions. ISAAQ also demonstrates its broad applicability, obtaining\nstate-of-the-art results in other demanding datasets.\n