Arthur Matei, Tim Hallyburton, Lukas Hennies, Christoph Rass and Gernot A. Fink
Proc. Int. Conf. on Document Analysis and Recognition, 2026, to appear.
Vienna, Austria
Extracting structured information from historical archival documents is challenging due to the scarce availability of labeled data and the oftentimes convoluted information on the documents. We build on the work of Wolf et al. (2025) to further investigate field extraction from CM/1, a dataset of historical WW II care and maintenance application forms, using vision-language models (PaliGemma and Donut) trained on as little as 1% of available annotations. To address data scarcity, we explore two strategies: cross-field pre-training on the same documents and synthetic document generation, where empty document templates are populated with synthesized handwriting. We show pre-training on synthetic documents yields small improvements, but cross-field pre-training shows greater improvements in low-data training. We further find that joint multi-field training slightly benefits larger models, while smaller models perform better when trained on single fields in low-data scenarios. Although our approach can be applied to any field in the forms, we exemplarily demonstrate practical relevance through nationality extraction, a field of high interest for historians in migration studies. Alongside this, we publish CM/1v2, an extended version of the dataset with more annotated fields, namely Nationality, Place of Birth, and Religion, covering both heads of family and family members. Notably, the Religion field poses a unique recognition challenge as it combines pre-printed checkboxes with free-text entries added by applicants. The dataset is available at github.com/rthrmt/cm1v2.