Unlocking Arabic Manuscripts: The Potential of Digital Humanities for Egypt’s Collections

Why One Manuscript Needs All The Others

The Case for a Unified Arabic Corpus

Maxim Romanov

September 27–28, 2026 · Cairo · The Evolution of Islamic Societies (c. 600–1600 CE), Universität Hamburg

Hosted & funded by

دليل المحاضرة بالعربية

ملخّصٌ مفصّل للمحاضرة بالعربية، مع شرحٍ موجز لكل شريحة.
امسحوا الرمز لقراءته.

maximromanov.github.io/slides/
2026-09-28-Cairo/companion-ar.html

How do we know what a word means?

ما لاعينرأت ولا أذن سمعت
ʿayn
eye“What no eye has seen and no ear has heard.”Ṣaḥīḥ al-Buḫārī, no. 3244
فيهاعينجارية
ʿayn
spring“In it is a flowing spring.”Qurʾān 88:12
فبعثعينعيناله من جهينة … فأتاه بخبر القوم
ʿayn
scout, spy“He sent a scout of his from Juhaynaŧ … and the man brought him news of the people.”al-Ṭabarī, Jāmiʿ al-bayān, on Q 8:7
ثمعينله السلطان محمد خان … كل يوم ثمانين درهما
ʿayyana
assigned“Sultan Meḥmed Ḫān then assigned him … eighty dirhams a day.”Ṭāšköprüzāde, al-Šaqāʾiq al-nuʿmāniyyaŧ, p. 94

We know it from what surrounds it.

The same framing at every scale

١A wordis understood through the words around it: its company restricts its meaning
٢A termis understood through the corpus: when does it appear across time and space? who uses it? and who doesn’t?
٣A bookis understood through other books: what other books it builds on? how does it fit among the rest?
٤A traditionis understood through the geographical and chronological layers of books: how does its language change? how do its topics shift?

The instrument

A Proxy to the Arabic Written Tradition

OpenITI as a chronological and geographical representation

OpenITI github.com/OpenITI

8,700+titles
3,300+authors
2.4 billionwords, all versions

Nigst, Lorenz, Maxim Romanov, Sarah Bowen Savant, Masoumeh Seydi, and Peter Verkinderen. OpenITI: A Machine-Readable Corpus of Islamicate Texts, Primary Version. Version 2025.1.9. Zenodo, 2026. doi.org/10.5281/zenodo.18613982

OpenITI, v. 2023.1.8 — “clean”

Our final subcorpus, based on version 2023.1.8, consisting solely of “clean” unique texts written up to 1400/1980:

  • – 2,783 unique authors (156 fewer than in the set above, or 94.69%)
  • – 7,058 unique texts (−407 fewer, or 94.55%)
  • – 927.37 million tokens before cleaning
  • – 868.81 million tokens after removal of most paraeditorial material and remaining “unclean” (93.68% of the uncleaned set)
Chronological distribution of the volume of the OpenITI subcorpus
The same chronological distribution, with the smallest period circled: even the thinnest slice, 1200–1300 AH, just 14 million clean words, is enough for many purposes

Case 1

A term in time

When did al-kutub al-sittaŧ come to be?

Case 1: al-kutub al-sittaŧ (الكتب الستة)

In “The Devil’s Delusions” (Talbīs Iblīs), Ibn al-Jawzī (d. 597/1201) uses lots of ḥadīṯs:


-   No isnāds for: Muslim, al-Buḫārī, and Abū Dāwūd


-   Full isnāds for: al-Tirmiḏī, Ibn Mājah, and al-Nasāʾī



No notion of al-kutub al-sittaŧ in the end of the 6th/12th century in Baġdād.

Google Books Ngram Viewer: telegraph, telephone, television
Mentions of al-kutub al-sittat over time
Authors mentioning al-kutub al-sittat by century and region

Case 2

A book among books

Text reuse and the modeling of text types

Case 2a

Text Reuse

By tracing at scale how later texts reuse earlier texts we can understand how each book fits into the tradition and how all the books form a network of interconnections

  • Among them was Abd Allãh b. Ḫāzim al-Sulamī, the governor of Ḫurāsān for Abd Allãh b. al-Zubayr. What was remarkable about him was that he was extremely brave and resourceful, yet he was terrified of mice. One day, while he was with ʿUbayd Allãh b. Ziyād, a white rat was brought before him and he was astonished. ʿUbayd Allãh said to ʿAbd Allãh, ‘O Abū Ṣāliḥ, have you ever seen anything more astonishing than this?’ And there was Abd Allãh, who had shrunk as if he were a chick and turned yellow as if he were a male locust. ʿUbayd Allãh then said, “Abū Ṣaliḥ disobeys the Merciful, takes lightly the authority, seizes the snake, walks toward the blooming lion, faces spears with his face and swords with his hands, and yet, he is overcome by a rat as you see. I testify that Allãh is capable of everything.

al-Ḏahabī and his Tārīḫ al-islām

Šams al-dīn al-Ḏahabī (d. 748/1348): historian, ḥadīṯ scholar, biographer.

Tārīḫ al-islām — a universal chronicle-cum-biographical collection covering seven centuries of Islamic history (1–700 AH / 622–1301 CE), organized decade by decade.

  • c. 30,000 biographies
  • c. 3 million words
  • Two modern editions: 50 vols. (Tadmurī) and 16 vols. (Maʿrūf)

Built largely from earlier sources — an ideal test case for text-reuse detection at scale.

Case 2b

Modeling Textual Typology

By modeling major textual types, we can identify types of unknown books and understand the structural composition of the complex books

Case 2b. Modeling Textual Typology

  1. [B] Biographical type (tarājim): mostly collections of biographies
  2. [E] Exegetical type: interpretations of the Qurʾān (tafsīr)
  3. [G] Graeco-Arabic (“Greek-esque”) type (yūnāniyyāt): translations from Greek and works on the development of Greek philosophy in Arabic (falsafat, predominantly Aristotelian)
  4. [H] Historical type (tawārīḫ): mainly chronicles
  5. [F] Legal type (fiqh): a broad variety of legal texts
  6. [L] Linguistic type: a broad variety of grammatical, rhetorical, and lexicographical texts (balāġat, naḥw, maʿājim luġawiyyat)
  7. [M] Medical type (ṭibb): mainly texts on Galenic medicine
  8. [P] Poetic type (šiʿr): mainly collections of poetry (dīwāns)
  9. [K] “Scholastic-theological” type: works of kalām
  10. [Ḥ] Tradition-based type: collections of ḥadīṯ

Testing the model: exegetical type (tafsīr)

The eight exceptions

  • Mafātīḥ al-ġayb by Faḫr al-dīn al-Rāzī (d. 606/1210) → classified [K] scholastic-theological: it is in fact a major work of kalām, and samples from it were used to train the [K] type
  • Ġarīb al-Qurʾān by Abū Bakr al-Sijistānī (d. 330/942) → classified [L] linguistic (still 26% exegetical signal): closer to a dictionary than to a traditional tafsīr
  • Six early ḥadīṯ-based interpretations — by Mujāhid b. Jabr, ʿAbd al-Razzāq al-Ṣanʿānī, Ibn Ḥakam al-Ḥibarī, al-Nasāʾī, al-Ṭabarī, and Furāt al-Kūfī — classified [Ḥ] Tradition-based: composed mainly of ḥadīṯ reports, so hardly a real mistake; three still carry a strong exegetical signal (28%, 42%, 49%)

73 of 81 texts (90.1%) landed correctly on the first pass; every exception is explicable.

Rolling Assessment: al-Šāfiʿī’s Kitāb al-Umm

Typological profile of al-Šāfiʿī's Umm across its length: the legal type dominates almost the whole book

Rolling typological assessment of Kitāb al-Umm of al-Šāfiʿī (d. 204/820) a single dominant type — [F] legal — running almost unbroken from beginning to end.

Rolling Assessment: Ibn al-Jawzī’s Talbīs Iblīs

Rolling typological assessment of Talbīs Iblīs of Ibn al-Jawzī (d. 597/1201) shows that individual chapters (separated by vertical black lines) usually contain text reflecting specific types.

Majmūʿāt Case

Value for manuscripts:

  1. When HTR starts working, we will need a mechanism to assess HTR-ed texts automatically: we can potentially identify type, author, exact text
  2. Particularly valuable for majmūʿāt — the most numerous kind of manuscripts, the most difficult to read and analyze

Case 3

A Tradition

language change and topic shifts through chronological layers of books

Case 3. Periodization

  • Preclassical   c. 622–750
  • Classical   750–1258
  • Postclassical   1258–1798
  • Modern   1798–
Preclassical600–750
Classical750–1258
Postclassical1258–1798
Modern1798–

Where do these boundaries come from? What if we ask the corpus?

Geschichte der arabischen Litteratur, 5 vols. (1896–1943)

Cambridge History of Arabic Literature, 6 vols. (1983–2006)

Encyclopedia of Arabic Language and Linguistics (Brill)

Two witnesses

First witness: Corpus

Chronological distribution of the volume of the OpenITI subcorpus

Two lenses on change

Keywords (TF–IDF): what texts are about
Style (MFV): how texts are written

  • Hundreds of parameter combinations
  • 3,000 random samples (MFV)
  • 72,000 dendrograms » 48 consensus trees

Do not trust a tree. Ask what survives across trees.

Contentkeywords / TF–IDF
Stylemost frequent vocabulary

Style

stylometry

  • distributions of frequencies of most frequent words (function words) are unique to individual:
  • authors
  • periods
Unrooted dendrogram of authors’ writings clustered by stylometric distance

Style

stylometry

  • distributions of frequencies of most frequent words (function words) are unique to individual:
  • authors
  • periods

Jurjī Zaydān’s writings
(1) 1891–1902
(2) 1903–1905
(3) 1906–1914

Dendrogram of Jurjī Zaydān’s writings in three periods
A single dendrogram of corpus samples by 50-year interval
A single dendrogram of 50-lunar-year samples of the corpus, colored by macroperiod. Hijrī dates!
Grids of small dendrogramsGrids of small dendrograms
72,000 dendrograms: hundreds of parameter combinations over 3,000 random samples.
Summary stripe diagrams, stylometric testsSummary stripe diagrams, keyword tests
Summary stripe diagrams: what survives across the trees.

Major periods

  • The early macroperiod (622–1200)
  • The middle macroperiod (1200–1540)
  • The late macroperiod (1540–1980)

Results of majority of 72,000 dendrograms:

3,000 stylometric samples (1,500 discrete; 1,500 overlapping)
12 distance metrics, 2 major clustering algorithms;
48 tf-idf keyword tests.

The middle macroperiod (c. 1200–1540)

  • Thematic profiles isolate it as a distinct period
  • Stylometric profiles often attach it to the early macroperiod
  • The language evolves slower than the written tradition: new distinct themes, but older language

Two clocks: themes shift faster than language.

A different macrohistory of the written tradition

  • Early: c. 622–1200
  • Middle: c. 1200–1540
  • Late: c. 1540–1980
  • Boundaries are transition zones, not event dates
Earlyc. 622–1200
Middlec. 1200–1540
Latec. 1540–1980

Two witnesses

Second witness:
Bio-Bibliographical Collection

Second witness: 8,800 authors

Ismāʿīl Bāšā al-Baġdādī (d. 1338/1919)

Hadiyyat al-ʿārifīn
c. 8,800 authors; 40,000+ titles
A bio-bibliographical model of the tradition

Independent of the corpus experiments

8,800authors
40,000+titles

c. 600–1200: an Iraqi-Iranian world

Map of the Islamic world with the flows of scholars in the early macroperiod, centred on Iraq and Iran
> 80% Iraq + Iran

c. 1200–1540: a Syrian-Egyptian world

Map of the Islamic world with the flows of scholars in the middle macroperiod, centred on Egypt and Syria
> 55% Egypt + Syria

c. 1540–1900: a new polycentric geography

Map of the Islamic world with the flows of scholars in the late macroperiod, with Anatolia and India prominent
≈ 30% Anatolia

Two transitions: 1200–1300 and 1500–1600

Map of the first major transition, with movements from Iraq and Iran westward and from al-Andalus eastward

Two transitions: 1200–1300 and 1500–1600

Map of the second major transition, with flows toward Anatolia and India

Two independent witnesses

CORPUS of TEXTS
lexical/thematic and stylometric change

REGISTER of AUTHORS
cultural geography over time

Different biases. Different assumptions.
Similar macro-boundaries.

Three long cultural regimes (macroperiods)

  1. The early Iraqi-Iranian  ·  c. 622–1200
  2. The middle Syrian-Egyptian  ·  c. 1200–1540
  3. The late Turco-Indian  ·  c. 1540–1980

(with discernable sub-periods inside each)

Periods are also places

  • Traditional periodization assumes universality
  • Corpus-driven, computational periodization demonstrates that language and textual production are aligned in time with changing cultural geography

Each period gets not just a name, but also a map.

What changes if we accept this model?

  • The Mamluk centuries become a coherent middle period, not the beginning of decline
  • c. 1540 emerges as a major structural break (essentially, the end of the Mamluk period); not 1258, the fall of Baġdād and of the ʿAbbāsids
  • Chronology & geography become part of the same explanation

The instrument, again

Back to the Corpus

Three case studies later — what can this instrument actually tell us?

But a corpus is also a historical source

  • Survival bias — what reached us
  • Editorial / digitization bias — what modern scholars chose to publish
  • Uneven genre and chronological coverage
  • Varying representation of different communities

Source criticism begins with the instrument.

Density plots of books by century in three libraries and the Hadiyyat al-ʿārifīn
Share of series total by century: Hadiyyat al-ʿārifīn, major libraries, and OpenITI unique texts

No text can be read alone.
Every text needs all the others.

No manuscript can be read alone.
Every manuscript needs all the others.

Cover of Digital Humanities for Arabic and Islamic Studies

In short, this book asks:

«What does it mean for our field when the entire library becomes a vademecum — something that, quite literally, “goes with me” everywhere?»

Chapter 1

Why we should even bother—a question that remains at the heart of every discussion about the digital.

Chapter 2

How we might go about harnessing the new digital opportunities within the field.

Chapter 3

What we can expect to accomplish:
1) trace term usage
2) model textual typology
3) chart linguistic evolution

Altogether, the book proposes a vision of digital humanities relevant to the field of Arabic and Islamic studies. It argues that this venture is not about cheating, but about keeping up with the rest of the world and ensuring that the future of the field remains in our own hands.

1 / 62

Keys

→ space PgDnnext slide
← PgUpprevious slide
Home / Endfirst / last
Ffullscreen
Ooverview of all slides
Sspeaker notes in a second window
Ttype tester: fonts, weights, size
+ / − / 0text larger / smaller / reset
?this help

Print (⌘P) gives one slide per page. Add #12 to the address to open a slide.