GQA-it: Italian question answering on image scene graphs

The recent breakthroughs in the field of deep learning have lead to state-of-the-art results in several Computer Vision and Natural Language Processing tasks such as Visual Question Answering (VQA). Nevertheless, the training requirements in cross-linguistic settings are not completely satisfying at the moment. The datasets suitable for training VQA systems for non-English languages are still not available, thus representing a significant barrier for most neural methods. This paper explores the possibility of acquiring in a semiautomatic fashion a large-scale dataset for VQA in Italian. It consists of more than 1 M question-answer pairs over 80k images, with a test set of 3,000 question-answer pairs manually validated. To the best of our knowledge, the models trained on this dataset represent the first attempt to approach VQA in Italian, with experimental results comparable with those obtained on the English original material.