Zusammenfassung:
A large literature in economic history uses digitized census data to study individual-level outcomes in history. Although many census records have been digitized manually, the process is extremely labor-intensive, and substantial material remains unprocessed in archives. Recent advances in machine learning offer the potential to automate large part of this work. We demonstrate an end-to-end digitization pipeline based on the transformer-based Donut model, trained on hand-annotated data and enhanced with image augmentation, to extract information from the 1955 Stockholm tax and census records. The resulting output attains high accuracy across multiple evaluation metrics.
