Work in Progress · On the bench

What we are building next.

Prototypes, studies and ideas at different stages. Here we show where they are going rather than how they work. If one of them speaks to you, as a partner, a pilot site or an investor, we would like to hear from you.

Featured · Ongoing · Digital Maktaba · Arabic-script collections

From the scan to the catalogue

من المسح الضوئي إلى الفهرس

Reading, understanding and cataloguing Arabic-script libraries.

One chain for a whole library. OCR models we train on Arabic-script title pages, then on the body of the text; vision-language and large language models that read the page and extract its knowledge and metadata; and cataloguing suggestions on top. The goal: tools to catalogue large volumes of digitised texts.

Every value in the proposed record says where it comes from: the page it was read on, or the open catalogue record that confirmed, corrected or completed it. The librarian has the last word.

The research behind it

The title page of Itḥāf al-akābir (Hyderabad, 1910) next to what was read on it and what the Library of Congress record confirms

From pixels to fields

The annotation tool: a title page with its OCR blocks boxed on the left, and the list of recognised segments with their tags on the right
Tag every block. In our annotation tool each OCR segment of a title page is boxed on the image and tagged with its role: title, creator, contributor, publisher, date, volume, noise. Thousands of pages annotated this way are the ground truth of the Digital Maktaba datasets.
Line segmentation of an Arabic printed page: each line in a box, with its baseline in orange
Find every line. Line segmentation of a printed Arabic page, the first step of the OCR we train: every line is boxed, every baseline traced.
The title page of Tārīkh madīnat al-salām with its blocks coloured by role: title, creator, relation, date, contributor, publisher, noise
Name every field. The same title page, coloured by role. The thin blue boxes are the OCR segments, the coloured ones the fields a cataloguer needs.

What it does

A book record with its outcome: edition identified in the Library of Congress, and the values corrected, added, confirmed or only read
See what changed. For every book, the values a catalogue corrected, the ones it added, the ones it confirmed, and those that only the page can vouch for.
Subject and Library of Congress classification added from the catalogue record
Gain what the page does not say. Subjects, classification and identifiers come in from the matching catalogue record, each with its link.
OCR · VLM · LLM كتاب الأغاني TITLEAUTHORPLACEYEAR Title pageTextMetadataRecord
One chain. From the title page to the text, from the text to the metadata, from the metadata to a record ready for review.

Research works from this project

This work grows out of the Digital Maktaba project, with the University of Modena and Reggio Emilia and the Giorgio La Pira Library in Palermo: datasets, benchmarks and tools published and peer-reviewed since 2021.

All the publications
  1. TPDL 2026 · forthcomingBenchmarking Extraction on Arabic Title Pages: From Traditional OCR to Vision Language Models
  2. TPDL 2025 · SpringerDM-LP: A Large Arabic Script Languages Dataset for OCR and Cataloguing Research ↗
  3. Umanistica DigitaleThe “Digital Maktaba LP”: Proposing a Comprehensive Dataset for Arabic Script OCR Title Pages in the Context of Digital Libraries and Religious Archives ↗
  4. IRCDL 2026 · CEUR-WS 4240The Codes Talk: Automated Extraction and Normalization of ISBNs for Metadata Integration ↗
  5. DigItalia 20Alfabeto arabo e OCR open source: un’analisi grafica e linguistica dei risultati di elaborazione per il recupero di dati catalografici ↗
  6. The Digital Turn in Religious Studies · V&RThe Digital Maktaba platform: a tool for semi-automatic document cataloguing and searching ↗

Next · Research prototype

Towards an agentic digital library

The next step is a library that works alongside its librarians. An agent reads the first pages of each volume, searches open catalogues and authority files on its own, proposes an enriched record and keeps the full trail of what it looked at and why. Cataloguing a large digitised collection becomes a matter of reviewing, not re-keying.

  1. Readsthe title page and the pages around it, in Arabic script, choosing the clearest copy of every detail.
  2. Searchesopen catalogues and authority files, such as the Library of Congress, Open Library and the Internet Archive.
  3. Proposesan enriched record in which every field cites its page or its source, for the librarian to validate.

Featured · Embryonic stage · Arabic, Italian, English

Arabo Matematico

العربية الرياضية

Embryonic stage · an idea taking shape, not yet a product

Every Arabic root that could exist, and the few that should not.

Arabo Matematico is a collaborative dictionary of the Arabic roots recorded by the lexicographers over time, across the different phases of the language's history. Set side by side, they show how the roots of Arabic have been used, how they have changed, and how the map of the roots in use has itself shifted from one age to the next. Arabo Matematico gives that change a visual form, and describes the potential of the language: the roots it has used, and the ones it still could.

It follows the study of the Arabic lexicon from the classical lexicographers to computational linguistics, where linguistic research and the digital humanities meet, looking for the link between mathematical computation and the roots of Arabic. It is built to be fully collaborative, and to serve researchers and students alike.

The main screen of Arabo Matematico: a root, its permutations and the classical lexicons

What you can do with it

The space of roots as matrices, one panel per first letterThe same space of roots as a cube in three dimensions
See the whole space. Every possible root at once: as matrices, one for each first letter, or as a cube you can turn in three dimensions.
The six permutations of the root k-t-b
Turn a root around. The permutations of every root, the way Khalīl ibn Aḥmad ordered his dictionary twelve centuries ago.
The verb kataba generated from its root and pattern
Play the patterns. Watch a root flow through the morphological patterns (awzān) that turn it into verbs, nouns and adjectives.

For research

For linguists and digital humanists: the roots collected by the lexicographers, phase by phase, in a single space that can be computed, compared and explored, and always traced back to its sources. A way to follow how the language grew, what it chose to use and what it left aside, how the study of the lexicon moved from the classical dictionaries to computational linguistics, and where mathematics and morphology meet.

For teaching

A root condenses a semantic field, and the morphological patterns turn it into words. Students who see this stop learning vocabulary word by word: they learn to build words by applying patterns to roots. Arabo Matematico makes that mechanism visible, in a tool simple enough for the classroom.

Featured · Evaluation study · Publishers, libraries, archives

(Re-)editor Assistant

Scanned books back to editable files, footnotes and all.

Ordinary OCR returns a wall of text and throws the scholarly apparatus away. The (Re-)editor Assistant turns scanned books, journals and archival pages into structured, editable documents in which every footnote is still tied to its marker: native Word footnotes, linked notes in PDF, Markdown and XML.

The (Re-)editor Assistant: a document goes in as a PDF, a TIFF or a folder of images

What it does

Output formats: Markdown, XML, Word and PDF, each with linked footnotes
Choose the formats. Word with native footnotes, PDF with linked notes, Markdown and XML, in a single download. Right-to-left scripts such as Arabic and Hebrew are laid out in their own direction.
Page preparation: split two-page spreads, trim edges and margins
Prepare the page. Split two-page spreads, trim running heads, footers and margins, alternate the gutter on facing pages, and check every crop on a live preview.
123123
Keep the apparatus. Footnote markers stay tied to their notes: click a marker in Word and jump to the note, and back.

For publishers

Reissue backlist and out-of-print titles without re-keying them: an editable Word file with working footnotes, ready for correction and new typesetting, with the critical apparatus of scholarly editions kept intact.

For libraries, archives and researchers

Heterogeneous and multilingual collections, single pages and two-page spreads, turned into searchable, quotable and reusable text.

On the bench

Five more projects between the lab and the market. Some will become services, some papers, most of them both.

  • 245 كتاب العين 880 Kitāb al-ʿAyn 020 978–•••–••••–• OPAC · SBN

    External validationCataloguing · Non-Latin scripts

    Non-Latin scripts in the national catalogue

    Non-Latin titles that can be found in their own script and in romanisation.

    Records in non-Latin scripts in the Italian national library catalogue, romanised by rule and cross-checked against the major international catalogues, so that a reader can find them in either script. The tests started with Arabic and are being extended to other scripts; the same method adapts to the catalogue, transliteration standard and scripts of each client.

  • HistoryLawLiteratureScience

    PilotCataloguing · AI

    A library that sorts itself

    Subjects for any collection, from the title page to the shelf.

    Assigning to books in any language the subject headings of the classification a library already uses, general or specialist, and handing back records ready for the catalogue. The first pilot runs on Arabic books with a research library's own scheme for Islamic studies; the approach adapts to any library, collection and classification.

  • قرآن 12 × IT 1882 → 2010 188219131914192919551967197919861989199319942010

    Published · Islam in History 2026Translation · AI · NLP

    Translation families of the Italian Qurʾān

    Twelve Italian Qurʾāns, 150 years of translation, one synopsis.

    A human-curated synopsis aligns, verse by verse, the complete Italian translations of the Qurʾān from 1882 to 2010: an expert-verified dataset that Italian has never had. On top of it, AI methods trace which translations form families and share stylistic and semantic affinities, where the older editions drift, and how single Arabic words such as raḥma, taqwā or fitna are rendered into Italian. The approach is modular and language-independent, ready for other languages and other sacred texts.

    Read the article in Islam in History, 4/1 (2026) ↗
  • ḤadīthTalmudPatres

    Corpus designAI evaluation · Religious studies

    When AI won't read scripture

    Measuring over-refusal on sacred texts.

    A scholar asking about a ḥadīth on war or a Talmudic passage on punishment is doing history, yet AI models often refuse or sanitise the answer. We are building a parallel corpus across the three monotheistic traditions to measure when, where and in which language this happens.

  • 14,19 · IUDITH τοὺς χιτῶναςGKτοὺςχιτῶναςαὐτῶνδιέρρηξανCconscid.×vestimentasuaIscider.×tunicassuasVscider.omnesvestimentasua

    Ground truth in progressCritical editions · OCR

    Vetus Latina, cell by cell

    The oldest Latin Bible, read the way its editors printed it.

    Synoptic critical editions stack Greek, several Latin lines and their variants on one page. We are turning them into aligned, structured data that keeps every editorial choice: in these pages an omission is not a blank.

Partner, pilot or invest.

Every project on this page is open to collaboration: a library that wants to be a pilot site, a publisher with a backlist, a foundation or an investor who sees where it can go. Tell us which one caught your eye.