← All posts

ProjectsMar 20266 min read

Five work packages for ancient languages: notes from ITSERR

Latin canon law, Arabic OCR, hagiographic legends, papal bulls and a video pipeline. Within ITSERR, Mem delivered five work packages on Latin and Arabic corpora for the University of Modena and Reggio Emilia.

By the Mem.srl team

The context

ITSERR (Italian Strengthening of the ESFRI RI Resilience) is a PNRR-funded project that strengthens the Italian node of RESILIENCE, the European research infrastructure for religious studies. Our contract with UNIMORE ran in March 2026 and covered two slots, one for Latin and one for Arabic.

Work packageCorpusWhat we delivered
WP3Corpus Juris CanoniciDataset correction and a layout-segmentation proof of concept
WP5Arabic printed booksImage Tagger, a web app for OCR annotation
WP6Videos · legendsPlorabunt pipeline · Sanctuaria OCR
WP7Papal bulls1,162 pages digitised and compared across editions

Latin: canon law and bulls

For the Corpus Juris Canonici we compared the existing dataset with Friedberg's critical edition (Leipzig, 1879–1881), annotating and correcting discrepancies across the Decretum Gratiani and the Liber Extra. In parallel we fine-tuned a YOLOv8m-seg model on 172 annotated pages of the Liber Extra, exported as ALTO XML, to recognise layout regions automatically.

The papal bulls asked for a different kind of care: 1,162 pages digitised, then a synoptic comparison of the transcriptions across the MGH, CCO and EFR editions.

To unite academic rigour with computational innovation: bridging scholarly tradition and digital futures.

Mem.srl, mission

Arabic: tagging OCR

OCR on Arabic print is only as good as what you can check. Image Tagger lets an annotator draw boxes on a page, link each Google Vision AI segment to its area and classify it, saving straight back to the source spreadsheets. More than 1,900 images were verified this way.

Legends and videos

Sanctuaria covers 40 hagiographic legends on the foundation of sanctuaries, from Italian printed volumes of the 16th to the early 18th century, transcribed with Transkribus and Google Gemini and then revised by hand. Plorabunt is an automated pipeline: from 227 source rows with a video link (187 unique URLs) it downloads, transcribes with Whisper and augments the metadata with Mistral.

What we take away

Every package paired a model with a human check: an edition to compare against, an annotator to confirm a box, a reviewer to read a legend. That pairing is what we bring to every project, and the code for this work is on our GitHub.

OCRLatinArabicLayout analysisLLM