Our study presents exploratory work using large language models (LLMs) to normalize Latin text created by HTR transcriptions of medieval manuscripts. One of the main challenges of using HTR for text generation in niche, pre-modern contexts is the limited number of open pre-trained models with which to do so, especially given how we aim to preserve scribal abbreviations and allographs and incorporate them  into model training data. In this short paper, working from the assumption that every HTR model makes theoretical assumptions about the kind of text that it aims to create, we compare transcriptions of medieval Latin manuscripts made in eScriptorium by three open HTR models. We chose open LLMs for purposes of transparency and reproducibility, in order to evaluate how they perform when prompted to normalize the transcriptions. Our work arises from more general attempts to align HTR-created transcriptions with different kinds of ground truth found in scholarly studies (prologues, titles, rubrics, incipits, explicits, etc).

Medieval texts are not fixed originals, but rather radically dynamic objects, and are continually reshaped through copying, scribal habit and use (Zumthor, 1970). At the core of the problem for automated normalization is that document-level orthography is not rule-based in a modern sense: variation reflects context, habit and transmission. Modern editions impose stability on manuscript traditions, suppressing variability and standardizing textual realities. While the rise of HTR in recent years has facilitated scaled access to archival documents, it can exacerbate the question of textual mouvance.

HTR models are research instruments designed with different approaches to transcription–varying from the most conservative that capture scribal data “as seen” in manuscript, to others that exhibit a high degree of normalization of text, similar to the work of a modern editor. We have experimented with three open HTR models in eScriptorium: first, CATMUS 1.6 model that regularizes certain features (u/v, i/j, and letter forms for r, d, and s) while maintaining medieval abbreviation distinctions, trained on a variety of genres (Pinche et al., 2023); second, TRIDISv2 model trained on diplomatic editions of documentary manuscript text, that normalizes letter forms and expands abbreviations (Torres Aguilar, 2024); and third, “FroLat”, trained on a macro-dataset of fully expanded transcriptions of Old French and Latin manuscripts and charters without expansion (Camps and Vidal-Gorène, 2023). On a sample of the prologue to the book of Judith, the following transcriptions were produced:

Table 1. A snippet from the biblical manuscript of Jerome’s prologue to the book of Judith, Aarau, Aargauer Kantonsbibliothek, MsWettF11, fol. 173v, along with the first seven lines of transcription using three HTR models: Catmus 1.6, TRIDIS v2 and FroLat. The normalized standard prologue (PL 29:37-39) is at far right.

Medieval scribes employed abbreviations where a single symbol or superscript mark could represent multiple letters and some scripts are complex to read. A variety of different issues arise in these short samples of transcription in Table 1. There is, of course, error in transcribed letters (the initial C in cuius, and the minims making up the letters t, r and n), but when combined with the inconsistent spacing and the splitting of words, the confusion in HTR transcriptions is more pronounced.

Whereas a scholarly apparatus for our test genre exists (Stegmüller, 1981), matching the incipits and explicits we have in Stegmüller to HTR created transcriptions of the sort found in Table 1 is, however, a non-trivial task. The challenge of highly variant text found in the manuscript sources combined with artefacts from the transcription process has led us to explore the use of LLMs to expand abbreviations and create a roughly “edited” text for string matching purposes. In our experience modern LLMs can perform context-aware expansion that accounts for grammatical case, word position, and semantic relationships. Whereas HTR results are often communicated in terms of character error rate (CER), even low CERs can produce a text that is difficult for an LLM to parse, raising important questions, about validation, transparency, and the role of AI-generated interpretations in scholarly workflows that we continue to explore.

We applied three open-source LMMs with some degree of exposure to Latin to our HTR-created transcriptions: Qwen3‑235B, GPT-OSS-120B, and MiniMax‑M2. Being open-source, they support reproducibility and broader accessibility in workflows for manuscript normalization. All three are designed for advanced reasoning and multilingual text understanding and have over 100 billion parameters, giving them strong capacity for handling noisy inputs and long contexts, factors which are crucial for analyzing HTR results. QWEN3 supports over 100 languages and dialects with strong capabilities for multilingual instruction following (QWEN Team, 2025). MiniMax‑M2 (MiniMax AI, 2025) and was included to experiment with the thinking mode, which breaks problems into smaller steps before executing the task, offering potential insights into model reasoning behavior and the effects of prompting strategies. Further, we selected GPT-OSS-120B (OpenAI, 2025) because it is specifically designed for precise instruction following and agentic text workflows.

In our first experiment we used a role playing prompting approach explaining the task of normalizing the Latin text, providing context for the kinds of issues it would encounter in the HTR-created transcription (Table 2). The LLM processing the three distinct HTR-created texts point to highly variable degrees of intervention. For the text created by the Catmus 1.6 HTR model that conserves many of the scribal features, QWEN3 is the most interventionist, resolving many of the Latin issues, including those due to spacing, providing the overall best text. MiniMax‑M2 makes less effort to remake the Latin and GPT-OSS does not provide a satisfactory result.

Table 2. The normalization prompt, and slight adjustment for the second experiment, for our transcriptions of four medieval Latin bibles.

For the Tridis transcription, the model trained on documentary genres to expand Latin the most, QWEN3 produces the best overall meaning of the text, especially staying close to the correct lemmas, however, with occasional drifting off topic. With all of Tridis’ expansions, this seems to lead all three models to change many of the words, cases and meanings in Latin. For the HTR transcription created with FroLat, the results are limited. Intriguing to the results are comments found in MiniMax 2’s thinking related to how it sees a need to interpolate expressions from the Latin bible.

On our best performing LLM transformations–Catmus 1.6, Tridis v2–we added a sentence of the target page produced by the HTR-model and the correct transcription, giving the model the idea of the combination of error and variance (Table 2). The second LLM experiment with Catmus 1.6 and Tridis v2 led to gains in terms of how the sentence spacing caused issues, but incorrect words from model memory persisted. The Minimax example again provided both in its thinking and changes in the text additional evidence of the positive impact of the additional ground truth.

Our study is limited to four manuscripts and three HTR models, but has resulted in interesting observations on model behavior worthy of future research. Notably, HTR models that preserve scribal behavior seem to result in a better normalization by LLMs. Prompt-engineering with additional information about HTR error seemed to provoke the most useful shift in LLM behavior. In short, normalizing text from HTR transcription has the potential to open up medieval manuscripts to human-in-the-loop NLP approaches. For example, an interactive review application for identifying scholarly ground truth in pre-modern texts combining fuzzy string comparison with embeddings (Bamman & Burns, 2020) would be possible to address alignment in the context of different scribal hands and regional abbreviation conventions.

Figure 1. Image of a potential review system for identifying scholarly ground truth in pre-modern texts.