Transcribing Historical Population Sources Written in Cyrillic: Methodological Challenges in Training HTR Models for Romanian Parish Registers in Transylvania
Transcribing Historical Population Sources Written in Cyrillic: Methodological Challenges in Training HTR Models for Romanian Parish Registers in Transylvania

Summary/Abstract: The large-scale digitization of archival holdings has created new opportunities for historical population research, but effective access to handwritten sources remains limited due to the absence of reliable automatic transcription tools. This paper presents the training of a Handwritten Text Recognition (HTR) model for the automatic transcription and transliteration of Romanian parish registers written in Cyrillic characters, a category of sources that constitutes a substantial yet difficult-to-access component of modern Romanian documentation. Focusing on Orthodox parish registers from Transylvania dating from the early nineteenth century, the paper combines a historical overview of Romanian Cyrillic writing with a methodological discussion of transliteration and transcription practices, followed by an empirical assessment of trials conducted using the Transkribus platform. The results obtained so far reveal high Character Error Rates and demonstrate that the standard HTR training workflow is insufficient for producing a functional automatic solution for this type of material. The main obstacles arise from the lack of orthographic standardization, the structural mismatch between the Cyrillic alphabet and the Romanian language, graphic polysemy, abbreviations, superscriptions, irregular spacing, and significant variation in handwriting. These difficulties show that transcription in this context cannot be fully automated and must instead be approached as a hybrid, semi automatic process that integrates HTR, rule-based transliteration, lexica validation, and sustained human intervention. By documenting both the progress achieved and the limitations encountered, this article contributes to ongoing debates in digital humanities and historical demography regarding the applicability of artificial intelligence to complex historical sources. It argues that, despite current constraints, even imperfect automatic transcriptions can significantly enhance accessibility and research efficiency, provided their use is methodologically transparent and critically informed.

• Details
• Contents