|
Visual Question Generation (VQG) is a novel task that has been proposed very recently. Though some VQG methods that are capable of generating diverse questions for images exist, they are not focused on generating question and answer pairs and their evaluation metrics are ambiguous. Furthermore, previous studies do not recognize generated questions as additional synthetic training data. In this paper, we propose a novel framework to generate synthetic question and answer pairs for Visual Question Answering (VQA) via adversarial training. Our experiment is evaluated on a semi-supervised setting of VQA and we compared our adversarial method with some implemented methods.
|
