What transcription approach HTR should learn? HTR, Normalization, and Editorial Tradition in Medieval Manuscript Studies

Extended abstract accepted for the International Workshop on Computational Paleography (IWCP 2026, 4th edition), 4 September 2026, Vienna, Austria, a workshop of ICDAR 2026, presented as a poster with a flash presentation. Published here for open access and citation. Workshop Organisers: Isabelle Marthot-Santaniello, Universität Basel, Basel, Switzerland Hussein Mohammed, Universität Hamburg, Hamburg, Germany Workshop page: IWCP 2026 : CSMC : University of Hamburg The growing use of Handwritten Text Recognition (HTR) tools radically changes how scholars engage with historical corpora, enabling them to process manuscripts on a much larger scale than ever before. As a natural side-effect, HTR also creates new methodological challenges. One of the major ones, which hasn't been discussed extensively so far, is, how HTR interacts with long-standing philological practices of various languages, influences them, or is in conflict with them. HTR can be theoretically trained using any type of transcription as ground truth: from hyperdiplomatic – where the transcription closely follows the original – to very interpretative ones, that, e.g., expand automatically abbreviations, normalize capitalization of letters, standardise orthography, or even change the script. Recently, we have seen a push towards (hyper)diplomatic HTR models. A common argument is that HTR models work best if they are trained on the data that closely correspond to what is present on the page, character to character, with only occasional small compromises (e.g., transcribing allographs "s" and "ſ" as the same "s" character). Additionally, it is being argued that such (hyper)diplomatic transcriptions may enable palaeographic inquiry, and that they follow best DH practices of the separation of various concerns. For some philologies, when working with historical manuscripts, the hyperdiplomatic approach is common (e.g., medieval German), and thus adoption of (hyper)diplomatic HTR systems does not cause methodological tensions. However, other philologies have developed over a long period of time various substantially different approaches to the transcription of historical manuscripts, which are also influenced by the research question, purpose of the transcription (e.g. preparation of an edition), and local conventions. Thus we may see anything from semi-diplomatic to very normalized transcriptions, or even highly sophisticated and research based silent conversions to modern orthographical system or intoa different script (so-called “transliteration”). Our contribution analyzes this methodological tension on the example of two distinctphilologies – medieval Latin (especially Late medieval) and medieval Czech – arguing that the contrast between them illuminates broader, underexamined assumptions embedded in the design of HTR ground truth, evaluation metrics, and post correction workflows.