Thibault Clérice

Ouip... Qui marche dans la page recherche classique

Of course a bug appeared at the last minute.
Of course...

Merci, et ce n'est que le début 💪

OH ! Pressing Enter works but the Search button does not...
Nobody caught that in preproduction... THANKS a thousand.

If you can do a screen capture, I'll be thankful :)

The main page does not have autocomplete normally 🤨

In which autocomplete field?

(11/🧵)
So here it is: CoMMA, a playground for philologists, linguists, and computational humanists alike.

Our goal is to make medieval textual variation computable.

💫 Explore → comma.inria.fr

#DigitalHumanities #ComputationalPhilology #HistoricalLinguistics #OpenScience #AI4Culture

comma-project/comma-jsonl · Datasets at Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

(10/🧵)

The corpus isn’t just readable 👁️ — it’s also fully downloadable!
Now hosted on @hf.co :

🧾 JSONL dataset → huggingface.co/datasets/com...
📂 More formats (ALTO, TEI, etc.) coming soon — we’re uploading the GBs as we speak.

(9/🧵)

And for historical semantics, the sheer scale changes everything.
Even rare words now have dozens or hundreds of attestations.

For instance, look at lasciuus — its semantic and morphological variants are suddenly visible across centuries. 🔍

(8/🧵)
From a historical linguistics or codicology angle, CoMMA already reveals fascinating trends 📈

💬 Latin and French show different peaks in the proportion of abbreviated tokens.
In Latin, the rate climbs to ≈ 9 % in the 14th century — a clear scribal difference with Old French !

comma-project/modernbert · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

(7/🧵)
🧠 Meet ModernBERT (yes, we see the irony — a medieval ModernBERT 😅)
Trained directly on CoMMA, and available here → huggingface.co/comma-projec...

We benchmarked it with masked-token prediction accuracy — and can’t wait to see how you’ll use it for historical or multilingual experiments!

(6/🧵)
So… what can we do with such a corpus? 🤔

We outlined three directions in the paper — here’s a taste of two:
1️⃣ Linguistic modeling & pretraining for medieval and early modern languages
2️⃣ Quantitative exploration of scribal practices, abbreviation systems & codicological patterns

(5/🧵)

The corpus wouldn’t exist without our partner and open repositories 🙌
A huge thanks to @biblissima.bsky.social for helping us identify relevant manuscripts.

Digitizations come primarily from:
📚 Arca (IRHT digitization database) → arca.irht.cnrs.fr
📖 Gallica at the BnF → gallica.bnf.fr

(4/🧵)
The CoMMA reading interface 🖥️ is built entirely on the Distributed Text Services (#DTS) protocol.
This means full interoperability — you can query, retrieve, and integrate CoMMA’s data just like any other DTS-compliant resource.
Open standards for open philology ✨

(3/🧵)
The dataset was built with the Kraken CATMuS 1.6.0 model (inria.hal.science/hal-04453952)

🪶 All transcriptions keep abbreviations.
📊 Sampled evaluation on 670 manuscripts → average CER = 9.7 %.

Some manuscripts are tricky, but overall, the ATR pipeline holds up well.

(2/🧵)
To put things in perspective:
⚙️ 2.2 bn “dirty” Latin tokens, compared to roughly 220 M clean Latin tokens known so far.
📜 Around 160 M “dirty” Old French tokens, versus the current ~10 M clean ones.

That’s a 10× jump in accessible data for premodern NLP and computational philology.

It's been brewing for months: @inriaparisnlp.bsky.social releases CoMMA (Corpus of Multilingual Medieval Archives) !

📚 2.5bn tokens of mostly Latin and French texts
🕰️ 800→1600 CE
📜 23k manuscripts
🖥️ 18k on the reading interface: comma.inria.fr
🔍 Paper: inria.hal.science/hal-05299220v1

(1/🧵)

🧵 Five years ago @yaelrice.bsky.social and I published this so that no one would have to reinvent the wheel of revealing why research like this is so misguided it defies sense. hyperallergic.com/604897/how-s...

but, because it is LONG and "detailed", it is given more weight.

If you are out there doing this, then why are you in the academy? This is a research conference, not an exam, and I am not your student to be 'corrected'.

TranscriboQuest Arabic Team : "The issue is layout recognition. So we worked on that."

Glosses in medieval manuscripts! Segmentation and transcription work

Ancient Greek team worked on Papyri from Herculanum, where all Segmentation and transcription have to be started from scratch

Catalog of books from a library in toulouse

Medical recipes from the 17th centuries on manuscripts written primarily by women.

Second presentation is a dataset for htr of 18th century Briton language manuscript

End of the TranscriboQuest 2025, funded by @biblissima.bsky.social and @atrium-eu.bsky.social and we are starting to present each team datasets.
First, medieval vernaculars with German, Swedish, Irish, Spanish

There is one person working on it AFAIK

It would be pretty hard as CATMuS does not provide markup-level information for paragraphs I think. (Sorry I was on vacation and cut all notificaitons ;) )