Synonym Replacement and Generation Enhancement for Document Augmentation

Document AI, or Document Intelligence pertains to the technology used for document comprehension and analysis. Given the multimodality of documents, the importance of multimodal learning cannot be overstated in the field of document intelligence research. Multimodal data augmentation, as a crucial aspect of multimodal learning, aims to enhance efficiency and accuracy of the model. This paper proposes a novel multimodal data augmentation approach called DADA, specifically designed for the document augmentation. The objective is to improve the expressive capabilities of multi-modal data and enhance the model robustness. To this end, we leverage synonym conversion and text-to-image generation techniques to augment text and image representations, respectively. Our proposed method is built upon a state-of-the-art document pre-training model. Experimental results on two small document datasets demonstrate that our approach can effectively enhance the performance of document comprehension and analysis models.