AI & Archaeology · 8th–16th centuries AD · Paris, Île-de-France, France
AI transcribes 32,763 medieval manuscripts, mostly Latin and Old French, in four months
The CoMMA corpus, built by researchers at France's Inria, makes manuscript texts searchable at scale, but hand checks found an error rate of 9.7%.
Not yet peer reviewed
Paris, France Researchers at Inria, France's national institute for digital science and technology, have used machine learning to transcribe 32,763 digitised medieval manuscripts. The job took four months. The result is CoMMA (Corpus of Multilingual Medieval Archives), a searchable collection of texts mostly in Old French and Latin.
The Inria computational humanities researchers worked with partners in France and Switzerland. Hand checks found an error rate of 9.7%. A paper describing the method and its limits is in preparation and has not yet been published.
Training the model on shared standards
Medieval handwriting is hard for machines to read. Most European vernaculars were still changing, and spelling was not standardised. New letterforms appeared, and pages are full of abbreviations, symbols and annotations.
When it came to transcribing medieval manuscripts, individual specialists went about it in their own way. But automating manuscript transcription requires machine learning, and for this you need standards.
Thibault Clérice, computational humanities researcher, ALMANACH team, Inria
The groundwork began in 2022 with a project called CATMuS. Ariane Pinche, a CNRS research fellow in medieval studies and digital humanities, launched it. She led the team with Clérice. They gathered 300 manuscripts that had already been fully or partly transcribed, about 200,000 lines in all.
The team brought these transcriptions into line with standards that respect medieval spelling and abbreviation. The documents date from the 8th to the 16th centuries. They cover about a dozen languages, mostly Old French and Latin, but also Spanish languages, Italian, Venetian and Dutch.
With this data the team trained a handwritten text recognition model. This is software that learns to turn images of handwriting into typed text. It relies on eScriptorium and Kraken, transcription tools developed at EPHE–Université PSL.
Inria describes the model as energy-efficient. It is built to recognise what is on the page instead of attempting broad language understanding. According to Inria's description, that kind of understanding can lead to confident but wrong guesses with medieval spelling, abbreviations or unusual hands.
From 300 manuscripts to 32,763
CoMMA began in 2024 with a larger team. Clérice managed the modelling and computing, joined by Benoît Sagot, head of the ALMANACH team. Hassen Aguili worked on the interface.
The manuscripts came through Biblissima+. This service gathers links to digitised manuscripts and catalogue data from many institutions, including the Bibliothèque nationale de France. The system has two parts. One algorithm finds and separates the elements of each page, such as main text, notes and illustrations. The CATMuS model then transcribes the text.
To test accuracy, the team checked three consecutive lines by hand in 670 manuscripts. They found an error rate of 9.7%. Some errors came from manuscripts older than the training material. Others came from hands that are hard to recognise, especially cursive scripts.
The team has not ruled out lowering the error rate further, but Clérice weighs the cost.
Doubling the processing time just to reduce the error rate by 1% wouldn't really be worth it.
Thibault Clérice, Inria
The corpus is much larger than earlier ones. The biggest previous corpus of manuscripts transcribed in Old French held 11 million pseudo-words (groups of characters). CoMMA has 516 million. For Latin, the figure has gone from 226 million words to 2.7 billion.
Elena Pierazzo, professor of digital humanities at the University of Tours, expects the corpus to help researchers use statistics to study how languages and dialects changed.
Having such a vast quantity of data that respects original spelling and abbreviations opens up all sorts of avenues for studying writing habits.
Elena Pierazzo, professor of digital humanities, University of Tours
According to the project's description, the corpus can now be used to train AI built to analyse old texts. Until now there was not enough data for that.
The team plans to add other languages by obtaining new texts from Biblissima+. Clérice said there was "no reason why Spanish or Italian languages, and the researchers studying them, can't take advantage of transcriptions from our model".
Sources
Spotted a mistake? Tell the editor about a spelling error or a fact that is wrong. We correct stories openly.
Readers' picks show on the front page.
Follow @AncientHeraldHQ on X for the day's finds as they come in.
Comments
House rules. Stay on the subject, be kind to other readers and say where your facts come from. No advertising and no personal attacks. The editor reads along and hides comments that break these rules.
Loading comments…