21/09/2026
Gutenberg’s First Phase: What We Built and What We Learned
When the Gutenberg idea began, the immediate question was straightforward: how can historical Arabic periodicals be made digitally accessible and searchable?
That was a logical starting point. The visible problem was access: scanned pages, fragmented files, and materials that were difficult to search computationally.
Working with a real corpus at scale changed the definition of the problem itself.
Digitization did not end the process.
It opened a sequence of deeper questions, and each layer we successfully built exposed another layer requiring different methods and tools.
After six installments in this documentation series, the experience can be summarized as a movement from:
**preserving the page toward reconstructing the knowledge carried by the page.**
This installment is not a declaration that the project is complete.
It is an attempt to document the first-phase outcome: what has actually been built, which problems proved most difficult, and how our understanding of Gutenberg changed from the original idea to the present.
1. The scale of the first phase
Gutenberg’s first phase includes seven Arabic periodicals:
Al-Risala, 1933–1953.
Al-Thaqafa, 1939–1953.
Al-Hilal, 1892–1953.
Al-Muqtataf, 1876–1952.
Al-Katib Al-Misri, 1945–1948.
Al-Zahra, 1924–1928.
Al-Kitab – Dar Al-Maaref, 1945–1953.
The total number of pages digitized and OCR-processed in this phase is:
**218,754 pages.**
This matters because the methodology has not been tested on a small, homogeneous collection.
The corpus spans decades, publication designs, editorial conventions, genres, authors, and intellectual domains.
But page count itself is not the objective.
The scanned page is only the beginning of the chain.
2. What came after OCR?
The first major outcome was a searchable textual layer.
But OCR quickly revealed that it could not reconstruct the periodical.
It extracts characters and words, but it does not automatically know:
where an article begins;
where it ends;
whether it continues several pages later;
whether a page contains multiple items;
whether a segment is an advertisement, marginal note, or index;
whether a heading is an article title or a recurring section;
or whether the item belongs to a series.
The project therefore moved from “text” toward “structure.”
3. Reconstructing the published item
We needed to treat the published item — article, poem, study, translation, or other contribution — as an independent unit.
But independence does not mean detachment from its source pages.
An item may consist of several segments distributed across non-consecutive pages.
A segment may begin halfway through a page and end before the bottom.
A single page may be shared by several items.
One of Gutenberg’s resulting principles became:
**the item is an independent entity, and every segment retains provenance to its source page.**
This allows a complete article to be presented without erasing its historical structure.
4. Why we needed a dedicated editor
Conventional text-editing tools were not designed for this problem.
We did not merely need to correct OCR characters.
We needed to:
mark item boundaries;
connect segments across pages;
identify titles and authors;
represent recurring sections and series;
separate articles from advertisements, marginalia, and indexes;
handle missing or recovered content;
preserve reading order while retaining source order;
and ensure that all of this structure survives export.
The Gutenberg Editor therefore evolved as a document-reconstruction environment rather than a simple OCR correction tool.
5. The source must not be erased
One of the most important scholarly and technical decisions was to preserve raw OCR as a source layer.
Correcting a text should not mean that the previous reading disappears.
Recovering material from another copy should not make that material appear as if it came from the original OCR source.
We therefore separated layers such as:
raw OCR;
editorial correction;
missing or recovered content;
human review;
and later publication/use layers.
This makes the origin of each state of the text traceable.
6. Provenance became part of data quality
Over time, provenance became one of the project’s central concerns.
The question was no longer simply:
Is the text correct?
It became:
Where did it come from?
Which page or image was its source?
Is it raw OCR or an editorial correction?
Was material added from another source?
Who reviewed it?
Which version was approved?
These questions make the data auditable and prevent the system from becoming a “clean text” repository with no visible history.
7. Export revealed errors that the interface could hide
One important lesson was that a successful interface does not guarantee a correct data model.
An article may look perfect inside the editor while export reveals that:
segments were flattened without page-level provenance;
an internal database identifier replaced the archival page reference;
advertisements or marginalia lost their type;
or section and series relationships were not carried forward.
This changed how we validate the editor.
The key question became:
**Can the origin of every exported segment be reconstructed from the data itself?**
If not, the archival structure is incomplete regardless of how correct the interface appears.
8. Human reviewers became part of the system design
Once review expanded to complete issues, it became clear that the platform could not be evaluated only from an engineering perspective.
It also had to be usable by human reviewers.
This led to simpler review workflows, including a local review environment in which reviewers could read an issue, see and modify structural markers, reorder article segments, add notes, preview the result, and return the reviewed data.
This reinforced another principle:
**a correct data model is not enough if the human workflow is unclear or unstable.**
9. Automated tests matter, but they are not the final judge
In a system like Gutenberg, a test suite can pass while the real editorial behavior remains wrong.
An interface can also appear correct while meaning is lost during export.
Our validation method therefore evolved to include:
automated tests;
live inspection in the editor;
review of real issues;
comparison with original pages;
and inspection of exported data.
These layers are necessary because a small software error in identity or linkage can become a serious scholarly error.
10. Once structure was solved, meaning became the next problem
After reconstructing items, another question emerged:
What is each article actually about?
We initially used free subject terms.
But a sample of 529 items produced 1,848 near-unique topical labels.
As the corpus expanded, thousands of additional expressions appeared.
This showed that uncontrolled indexing would not produce a usable system for browsing and discovery at scale.
11. Not every “keyword” was a subject
Reviewing the free-term lists showed that many entries were actually:
persons;
places;
works;
organizations;
events;
historical periods;
literary forms;
or motifs.
For example:
“poem” is a Form/Genre, not a subject.
“Egypt” may be a Place.
“al-Aqqad” is a Person.
“elegy” may be a Motif in a creative text, or a Subject in a critical study about elegy.
This distinction reshaped the indexing model.
12. From keywords to knowledge organization
This led to separate knowledge layers.
GAPT:
the Gutenberg Arabic Periodicals Taxonomy.
Its role is broad hierarchical classification and browsing.
ART:
the Al-Risala Thesaurus.
Its role is to control subject descriptors, alternative terms, and semantic relationships.
Alongside them:
Authorities for persons, places, organizations, works, and events.
Form/Genre for document and literary forms.
Motifs for literary themes and purposes.
This prevents every type of knowledge from being forced into one subject hierarchy.
13. Why we did not copy an external vocabulary wholesale
We used Arabic subject-heading and thesaurus references.
But the working principle became:
**the corpus determines what concepts we need; reference sources help us name and control them.**
We do not import thousands of headings merely because they exist in an authority source.
We do not automatically convert every “see also” reference into an RT relationship.
We do not assume every compound heading belongs in the subject thesaurus.
Each candidate must be reviewed within the logic of the corpus.
14. Frequency does not create a concept
Term frequency became very useful for prioritization.
But a word occurring dozens of times does not automatically deserve an independent subject descriptor.
It may be:
a place;
a person;
a form;
an alternative label;
or a concept already represented elsewhere.
Frequency therefore became:
**a diagnostic and prioritization tool, not a classification authority.**
15. Stable identifiers matter more than labels
As ART developed, stable concept identifiers became essential.
Labels can change.
Concepts can move.
Relationships can be revised.
Two descriptors may eventually be merged.
If thousands of records are linked only to a text label, every terminological change becomes a data problem.
A stable identifier preserves concept identity across versions.
16. SKOS gave us interoperability
Representing ART in SKOS was an important milestone.
It allows machine-readable representation of:
concepts;
preferred labels;
alternative labels;
broader/narrower relations;
related terms;
collections;
and other semantic structure.
This opens a path toward APIs, interoperable research tools, and future linking with other knowledge systems.
But another lesson followed:
**technically valid SKOS does not guarantee a scholarly valid thesaurus.**
17. Why ART remains prepublication
The current ART prepublication state includes:
149 active subject descriptors.
12 concept groups.
562 RT relationships.
15 broader/narrower relationships.
40 UF forms across 34 descriptors.
And valid, reparsable SKOS export.
However, a corpus-coverage audit revealed 277 recurring subject terms appearing at least five times that still lack a final editorial decision.
Of these, 89 occur ten times or more.
It would have been easy to ignore that gap and announce a “final” release.
We chose not to.
ART remains prepublication until real-world validation is completed.
18. The pilot is not ceremonial
The next ART stage is a pilot on real Al-Risala materials.
Two independent reviewers will index the same items.
Neither sees the other’s decisions.
We will examine:
coverage;
agreement on primary subject;
missing descriptors;
scope-note problems;
UF needs;
and confusion between Subject, Motif, Form, and Authority.
The corpus itself will therefore test the thesaurus.
19. Where AI helped — and where it stops
Automation and AI have been useful across many project stages:
OCR;
structural suggestions;
entity extraction;
classification assistance;
vocabulary review;
relationship suggestions;
and corpus analysis.
But one rule remained stable:
**a model produces a proposal or a version, not final archival truth.**
A model can be highly confident and still wrong.
Automated output must therefore remain attributable, reviewable, and revisable.
The published state is curated and approved.
20. The most difficult scholarly problems
The major scholarly challenges can be grouped into several categories:
**A. Unit of description**
Do we work at page, item, or issue level?
**B. Item boundaries**
How do we identify beginnings, endings, and continuations?
**C. Source linkage**
How do we preserve the relationship to the original page?
**D. Text layers**
How do we correct without erasing OCR?
**E. Identity**
How do we distinguish author from person discussed?
**F. Aboutness**
What is the item actually about?
**G. Dimension separation**
What is Subject, Form, Motif, or Authority?
**H. Terminological control**
When are two expressions genuinely equivalent?
**I. Semantic relationships**
When is a concept broader, narrower, or related?
**J. Coverage**
When does the thesaurus sufficiently represent the corpus?
None of these questions had a single ready-made answer.
They required repeated decisions and testing against real materials.
21. The most difficult technical problems
Technical challenges were inseparable from scholarly ones.
Among the most important:
variable-quality Arabic OCR;
multi-column page layouts;
articles spanning non-consecutive pages;
multiple items on a single page;
preserving source identifiers;
representing editable markers without modifying raw text;
reordering segments inside working copies;
exporting provenance at segment level;
handling recovered content;
avoiding duplicated logic across interface, import, and export;
and building workflows that human reviewers can actually use.
A seemingly small engineering choice can therefore have long-term archival consequences.
22. How Gutenberg’s definition changed
At the beginning, Gutenberg might have been described as:
“a project to digitize Arabic periodicals and make them searchable.”
That description is no longer sufficient.
A more accurate description today is:
**Research Infrastructure for Arabic Periodical Heritage.**
An environment connecting:
digital preservation;
text;
editorial structure;
source provenance;
metadata;
entities;
deep indexing;
and knowledge organization.
The project moved from focusing on the “digital document” toward the relationship between the document and the knowledge that can be responsibly derived from it.
23. What is not finished yet?
Good documentation should record unfinished work as well as achievements.
ART is still prepublication.
The thesaurus pilot has not yet been completed.
Coverage gaps still require review.
Some corpus layers will require further editorial review.
Editorial workflows will continue to evolve as they are used.
Expansion into additional periodical collections should follow only after first-phase tools and methods have demonstrated sufficient stability.
These are not signs of failure.
They are the boundaries of the current phase.
24. What does “first phase” mean?
The seven periodicals do not define Gutenberg’s final scope.
They form the foundation corpus.
The purpose of the first phase is to use a sufficiently large and complex body of materials to expose real problems in OCR, structure, export, indexing, and knowledge organization.
The goal is to develop methods and tools that can later be reused when the project expands.
The value of the first phase therefore lies not only in its 218,754 pages.
It also lies in the testing environment those pages created for the entire system.
25. What is the real measure of success?
Page count matters.
Item count matters.
Descriptor count matters.
Relationship count matters.
But none is sufficient on its own.
A better measure consists of questions such as:
Can the researcher reach the item?
Can they identify its source?
Can they return to the original page?
Can raw OCR be distinguished from correction?
Is the item structurally complete?
Are its subjects controlled?
Are persons, places, works, and forms kept distinct?
Can another system read the data?
Can another reviewer understand and reproduce the editorial logic?
When these conditions are met, digitization begins to become research infrastructure.
26. The larger lesson
Perhaps the most important lesson is that the best decisions were not always those that added the most features.
Often, the right decision was:
not to erase raw text;
not to import every external term;
not to add a relationship without evidence;
not to declare a final version too early;
not to trust automated tests alone;
and not to link thousands of items before validating the thesaurus.
Quality sometimes came from **refusing to move too quickly** as much as from building new components.
27. From here to the next stage
The next work is not a separate project.
It is the continuation of what has already been built.
First:
validate ART on real materials.
Then:
review the gaps exposed by the pilot.
Then:
publish a stable release suitable for broader linking.
In parallel:
continue corpus review and improve researcher-facing access and tools.
Expansion to additional periodicals becomes safer once the first-phase methodology has demonstrated that it can operate consistently.
Conclusion
Gutenberg began with a question about access to historical Arabic periodical pages.
The work led to deeper questions:
How do we preserve the source?
How do we reconstruct the article?
How do we describe it?
How do we distinguish subject from form and entity?
How do we control terminology?
How do we represent semantic relationships?
And how do we prove that a system works before declaring it complete?
The outcome so far is not merely 218,754 digital pages.
It is a set of layers, tools, and methodological rules forming the foundation of a research infrastructure for Arabic periodicals.
Perhaps this is the largest change in our understanding of the project:
**the goal is no longer simply to put heritage online, but to preserve its interpretability, traceability, and analytical value as it becomes data.**
With this installment, the first foundational documentation series reaches a natural closing point.
The next documentation series can move from “how was Gutenberg built?” toward more specialized topics: editor design, ART and GAPT methodology, Arabic OCR quality, indexing models, researcher tools, and interoperable data.
Send a message to learn more
