Synthetic document generation for the task of visual document understanding

Solving the problem of document analysis using machine learning methods requires a large amount of labeled data. Such data is not always available, and if available, it only covers certain types of documents.
In this paper, we present a method for creating synthetic data that allows creating documents of any type by pre-defining the document components. By changing the arrangement of document components, text content, and visual elements using configurations, we create diverse and realistic datasets that mimic real documents. This method addresses the problem of the lack of labeled datasets and offers a flexible solution to improve the results of a machine learning model.