InsureTVQA: A Progressive Text-Visual Question Answering Method for Insurance Domain

Visual question answering (VQA) is a multimodal task which answer a question related to an image. Existing VQA methods tend to focus on the target object on the visual level and ignore the textual information in the image, in the actual insurance domain, what needs to be focused on should be the textual content in the image and comprehends the text semantics. This paper proposes a progressive method InsureTVQA, which integrates DBNet text detection model, CRNN text recognition model, and RoBERTa-wwm machine reading comprehension model. In addition to this, the paper proposes a text box reorder algorithm and a text relevance filtering algorithm. Experiments are conducted on the AIWIN2021 Insurance Text Visual Cognitive Question Competition dataset, and the results show that the method can effectively learn the textual contextual information of insurance domain images and complete text visual question answering.