Skip to content

Reading edition

Dormant Digital Collections, AI Triage

Frédérick Madore

Deutsche Gesellschaft für Asienforschung · 6 May 2026

17 slides

Contents

Slide 1

Deutsche Gesellschaft für Asienforschung · Roundtable: DH & KI in den Regionalwissenschaften

Dormant Digital Collections, AI Triage

Lessons from the Islam West Africa Collection and beyond

Frédérick Madore Data Curator Africa Multiple Cluster of Excellence

Slide 2

Who I am

A historian, working with AI

  • Data Curator, Cluster of Excellence "Africa Multiple", University of Bayreuth
  • Historian of Islam in Francophone West Africa since the 1960s — youth, women, politics, education
  • Developer of the Islam West Africa Collection (IWAC) since 2021
  • Coordinator, Digital Humanities and AI in African Studies
  • Edited volume forthcoming, Bielefeld University Press (2027)
Book cover: Religious Activism on Campuses in Togo and Benin by Frédérick Madore (De Gruyter)
Monograph · 2025
Book cover: Religiosity on University Campuses in Africa, edited by Sounaye and Madore (LIT)
Co-edited · 2023
Blog: Charting New Territory — Digital Humanities und Künstliche Intelligenz in den African Studies, by Frédérick Madore and Vincent Hiribarren
DH & AI initiative

Slide 3

The allure of the archive

From the reading room to the "scanning party"

The reading room
Book cover: The Allure of the Archives by Arlette Farge — a magnifying glass over an old handwritten manuscript

Farge (1989): the slow, tactile encounter — "touching the real".

The scanning party
An unsorted backroom: stacks of old newspapers and bound volumes piled on tables, chairs and a sofa, with a scanner in the foreground — awaiting digitisation

An unsorted backroom, photographed in three days.

The shift Thousands of JPEGs and hours of recordings, half-named, untouched for years — affect moves from slow intimacy to frantic accumulation.

Slide 4

From backlog to triage

One historian, no team

  • Fieldwork across West Africa, plus libraries in Berlin, Paris, Washington
  • Public side: the open-access IWAC — islam.zmo.de
  • Cataloguing it all by hand is impossible
The question I put to AI Not "read this for me" — but where is close reading worth my time?
A digital hoarder, in numbers
30,000+items collected (Zotero)
~15,000still unprocessed
14,500+open access (IWAC)
1historian, no team

Slide 5

The shape of the talk

Six uses of AI in the research workflow

1Text extractionOCR and HTR with multimodal models
2Audio & video transcriptionincluding under-resourced languages
3Entity & topic extractionwho, where, what, about what
4"No-code" visualisationthrough AI-assisted coding
5First-pass analysis at scalesentiment, classification, triage
6NotebookLMchat with your own sources

Slide 6

Use 1 · Text extraction

Multimodal LLMs read the page like a person

A multi-column Ivorian newspaper article, 'Les imams se battent', overlaid with red numbered markers showing the reading order that traditional OCR scrambles — labelled 'Layout recognition failure'

Layout recognition failure: traditional OCR scrambles a multi-column page.

  • Traditional OCR breaks on multi-column papers, tables, handwriting, quick phone scans
  • Even a 5–10% error rate matters: one typo in a name hides a document
  • "Ouédraogo" → "Oueclaoqo": the source is there, but no one will find it
  • Multimodal AI handles it in one pass: printed & handwritten, French & Arabic, messy layout
Strongest so far Google Gemini 3 Pro — the best model I have tested for this.

Slide 7

Use 1 · in practice

Reading printed Ewé

A 1976 page of the Togolese newspaper Togo Presse, printed in the Ewé language with full diacritics

Togo Presse, 31 July 1976.

Extracted text · Ewé
MIAFE ŊUTIFAFAFIA EYADEMA KPE TA KPLE MƆSLEMTƆWO FE ƉEKAWƆWƆ DƆDZIKPƆHA LA

Mia nɔvi Mɔslem hamea ngɔnɔlawo se nya xɔasiwo tso Togo yeyea fofo nu

«Woayra amesiwo léa avu la, elabena woayɔwo be Mawuviwo»

Enye Togo yeyea fofo avafiawodzifia Gnassingbe Eyadema fe didi vevi be amesiame nanɔ dzidzɔ kple ŋutifafa me lle yefe Denyigba la katã dzi, Dzre, fuéle kple mamãwo manɔ amewo kple habɔbɔwo domo o. Eyaŋuti dukplɔla Eyadema anukwaretɔ la lɔna ɖɔɖɔɖowo wɔwɔ edziedzi le dukɔmeviwo dome ɖo.

Ete ŋu dze edzi be miafe tatɔ nutefewɔla la do go kple mia nɔvi Mɔslem Hame fe amegãwo nyitsɔ le RPT fea me hena wodome nyawo sese,

Tototɔ kple dzre ɖo Mɔslem Hamea fe dɔdzikpɔha la me etefe didi. Ale mamã kple fuléle ɖo wo dome.

Le wofe anyinɔnɔa kple numeɖeɖeawo me na To o yeyea fofo vɔ megbe la, miafe ŋutifafa anukwaretɔ Eyadema ɖo asi Hamea fe dɔdzikpɔha la dzi elabenà ŋutif fa kple lɔlɔ̃ mele wo dome o

Eɖe nu me na Xɔsetɔawo be ɖekawɔwɔ kple ŋutifafa dim yele le yefe anyigba la dzi..,

Mlɔeba la dukplɔla la ɖo dɔléɖeasihaa ɖe, si me ame 17 le la na mia nɔviawo bena woadzra Mɔslemtɔwo fe takpekpe gã. si wɔ ge woala la le Lome le Dasiãmime 4 lia dzi la fe dɔwɔwɔwo ɖo,

Avafiawo Menveyinoyu Dzafalo, dũdɔnunɔla Lãmesẽ kple Hadomegbɛnɔnya-wogbɔkpɔla kple Ayaovi Asila, Togo tuakɔwo fe tatɔ kpeɖeŋutɔ hekpe ɖe dukplɔla fe dɔwɔfe dzikpɔla afɛtɔ Kpotivi Têvi Dzidzagbe Lacle ŋu woenɔ tatɔ Eyadema ŋu le Mɔslem Hamea fe nyawodɔdrɔ̃ me. Eye afɛtɔ Mama Fuseni kple Kasim Mensa, siwo nye Hamea fe bubu 'mewo la hã nɔ Takpekpea me nyitsɔ ma,

Slide 8

Use 1 · in practice

Reading cursive French

A handwritten 1991 letter in French from Muslim associations in Burkina Faso addressed to the Minister of Information and Culture

Handwritten letter, Burkina Faso, 1991.

Transcribed text · French

3 H 1 a

Des musulmans du Burkina Faso,

  • Communauté Musulmane
  • Mouvement Sunite
  • Association de la Tidjania.

au Camarade Ministre de l'Information et de la Culture, Ouagadougou

Nous venons, par cette mise en garde, vous entretenir d'un sujet qui nous tourmente depuis des années. La goutte d'eau fait déborder le vase dit-on et c'est la raison pour laquelle nous pensons qu'il est temps de vous sensibiliser, de vous mettre en garde.

Mardi 16 avril 1991. Fête de fin du jeûne musulman. Journée chômée et payée en raison d'une loi de l'État.

Mardi 25 décembre 1990 Fête de la nativité, fête chrétienne Journée chômée et payée en raison d'une loi de l'État.

Il y a là une coïncidence heureuse mais facheuse à l'analyse.

Mardi 25 Decembre 1990. Jour de la semaine coïncidant avec une fête ; cette fête est chrétienne, le ministère décide : les mass media, patrimoine de tous les Burkinabè, sont à la disposition des chrétiens. La télévision qui d'ordinaire n'émet que les seuls samedis et dimanches dans la journée s'organise et dès 12h30, elle émet et diffuse le message chretien jusqu'à la nuit.

Mardi 16 avril 1991 Egalement jour de la semaine coïncidant avec une fête. Cette fête est musulmane. Le ministère occulte : la direction de l'information fait comme bon lui semble. La Télé

Highlighted: the writer's own spelling slips, kept verbatim — the model didn't silently correct them.

Slide 9

Use 2 · Transcription

Audio and video transcription

  • The same multimodal logic extends to audio and video
  • Transcripts return with timestamps and speaker labels
  • IWAC corpus: ~155 hours of Nigerian preachers, mostly Hausa & Arabic
  • Early tests in Hausa, Arabic, Kurmanji Kurdish — promising as a first draft
A video still from the IWAC audiovisual corpus: a preacher delivering a sermon titled Demokuradiyyah A timestamped, speaker-labelled transcript of a Hausa and Arabic sermon, generated with Google gemini-3-pro
A sermon and its machine transcript — timestamps, speaker labels.

Slide 10

A caution

Cleaner than the source

  • AI-OCR and audio-to-text open previously unusable material
  • The model silently modernises spellings, standardises toponyms, smooths regional variants
  • It can also hallucinate: invent text that was never there
  • The failure mode is harder to spot: a fluent paragraph with invented detail
The discipline Keep the original image alongside the text. Verify against the scan before you quote.

Slide 11

Use 3 · Extraction at scale

Entities and topics: who, where, what, about what

A network graph from the IWAC dashboard linking a central figure to the people, organisations and places that co-occur with him across the corpus

Entities linked across the corpus.

  • Manual keyword tagging: unworkable across thousands of documents
  • Standard DH tools trained on Western text — they miss local names
  • OCR noise compounds the problem
  • LLMs extract entities and, given the right prompt, topics
A feasible query, now "Find me every document that discusses secularism or the pilgrimage."

Slide 12

Use 4 · "No-code" visualisation

From text repository to knowledge base

An explorable layer over the documents — built without hand-written code.

A world map from the IWAC dashboard with article counts plotted by location, concentrated in West Africa
Mapped · articles by place
A network graph linking people, organisations and places that co-occur across the IWAC corpus
Linked · who appears with whom
A heat-map matrix showing how themes such as radicalisation, extremism, terrorism and jihadism co-occur across articles
Co-occurrence · themes that travel together
A word cloud of the most frequent terms across the corpus, led by musulmans, président and pays
Topics · the corpus at a glance

Slide 13

Use 5 · First-pass analysis

Sentiment analysis, without writing code

The IWAC Sentiment Analysis dashboard comparing three AI models across 12,287 articles, with discrepancy counts and breakdowns by polarity and subjectivity

Three models compared, article by article. fmadore.github.io/IWAC-sentiment-analysis ↗

  • 12,000+ articles, three LLMs, three axes: polarity towards Islam, subjectivity, centrality of religion
  • ~€35 and 24 hours of batch processing — months of manual coding avoided
  • Triage, not full reading; model disagreement is itself a signal
A pattern I'd have missed Positive portrayals of Muslim communities inside Sahelian terrorism coverage.

Slide 14

Use 6 · Chat with your sources

NotebookLM: the lowest-friction entry point

The NotebookLM interface loaded with IWAC newspaper articles: uploaded sources on the left, an answer with inline citations in the centre

NotebookLM over the IWAC newspaper corpus.

  • Only knows what you upload; cites the exact passage for every answer
  • Your sources are not used to train models
  • Free tier: 50 sources per notebook (PDF, web, YouTube, audio)
  • Multilingual: ask in your language, it searches sources in any of 80+

Slide 15

Hands-on

The same workflows, in your browser

1OCR & HTRPDFs and images, printed or handwritten
2Audio & video transcriptioninterviews and lectures, with speaker labels
3Summaries & keywordsbatch a whole spreadsheet of texts
  • Google Colab notebooks — nothing to install
  • A free Gemini API key; results save to your Drive
  • Open to anyone; written for researchers, not developers

Slide 16

The stakes

Western tools, non-Western sources

The costs
  • Training data overwhelmingly Western — concepts read through European lenses
  • Labour: Kenyan annotators paid ~$5/day; similar conditions elsewhere
  • Infrastructure: strongest models in US firms; only China credible as an alternative
  • Opaque and non-reproducible: the same input can yield different answers
What it makes possible

Research at scales, and with documents, not previously feasible — degraded scans, under-resourced languages, corpora too large to read.

Held in tension Use the tools, and disclose what they cost.

Slide 17

In closing

The scans on our hard drives are the raw material