Classical and Quranic Arabic NLP: Annotation for Religious and Heritage AI

Why Classical Arabic Is a Different NLP Problem From Modern Arabic

Modern Standard Arabic (MSA) — the formal written Arabic used in news, government, and education — descends from Classical Arabic but differs from it substantially. Classical Arabic is the language of the Quran, of pre-Islamic poetry, of the great classical Islamic scholars (Ibn Sina, Al-Ghazali, Ibn Khaldun), of medieval Islamic jurisprudence, and of a rich tradition of history, philosophy, and literature produced over fourteen centuries.

The morphological gap is significant. Classical Arabic retains the full case-marking system (i'rab) that MSA uses in formal contexts but drops informally. Classical verb forms include patterns rare in MSA. The jussive and energetic moods, common in Quranic verse and classical prose, appear infrequently in modern corpora. An NLP model trained on contemporary news Arabic lacks the morphological coverage to parse classical texts accurately.

The vocabulary gap is equally substantial. Classical texts use words and semantic senses that do not appear in modern Arabic — or that appear in modern Arabic with shifted meanings. A classical term for “scholar” may differ from the contemporary term. Legal vocabulary in classical fiqh uses technical definitions that were codified over centuries of jurisprudential debate. An NLP model that has not been trained on classical vocabulary cannot perform NER or topic classification on these texts reliably.

According to research from the Arabic NLP lab at NYU Abu Dhabi (CAMeL Lab), leading Arabic NLP models — including AraBERT and CAMeLBERT — achieve only 45–62% F1 on Classical Arabic morphological tagging tasks in zero-shot conditions, compared to 85–92% F1 on equivalent MSA tasks. The gap is not architectural — it is a training data gap.

The Islamic AI market amplifies this challenge. Across Saudi Arabia, UAE, Qatar, and Malaysia, a wave of Islamic fintech, Islamic knowledge platforms, Quran recitation apps, and Islamic legal AI systems is creating demand for NLP on Classical and Quranic Arabic text at scale. Saudi Arabia's Vision 2030 explicitly includes Islamic heritage digitisation as a cultural objective. The King Abdulaziz Public Library has digitised over 4 million manuscripts. These digitised collections become AI training opportunities only when correctly annotated.

What Quranic Arabic Annotation Involves

The Quran is the most-studied text in Arabic NLP, and the most consequential to annotate correctly. Several open annotation datasets exist, most notably the Quranic Arabic Corpus from the University of Leeds, which provides morphological annotations (part-of-speech, root, stem, features) for every word in the Quran. This corpus is a foundational resource for Quranic NLP research.

However, commercial Quranic AI applications require annotation beyond what open datasets provide:

• Tajweed annotation: For Quran recitation AI (increasingly deployed in apps across the GCC and Southeast Asia), annotation must capture the pronunciation rules of Tajweed — madd elongation, ghunna nasalisation, idgham assimilation, qalqalah echo. This requires annotators who are themselves certified in Tajweed, not just Arabic speakers.
• Thematic classification: Verses grouped by theme (tawhid, ethics, legal rulings, eschatology, historical narrative) for Islamic knowledge platform search. Theological correctness matters here — a misclassification is not merely a precision/recall error but a potentially significant error in religious context.
• Ahkam annotation: Identification of legal rulings (wajib/obligatory, haram/forbidden, mustahabb/recommended, makruh/disliked, mubah/permissible) in verses that carry fiqh implications. This is the highest-expertise annotation task in Islamic AI.
• Cross-reference alignment: Linking Quranic verses to relevant Hadith, tafsir commentary, and fiqh rulings for knowledge graph construction. Requires both Classical Arabic competency and Islamic science domain knowledge.

A critical principle for all Quranic annotation: the annotation must be reviewed by scholars with appropriate religious credentials for any task that involves interpretation of meaning. This is not just a quality requirement — it is a cultural requirement that affects whether the resulting product will be accepted by its intended audience.

Hadith and Classical Islamic Text Annotation

Hadith literature — the collected sayings and actions of the Prophet Muhammad (peace be upon him) as recorded by his companions — is the second major corpus for Islamic AI annotation. The six canonical Hadith collections (Sahih Al-Bukhari, Sahih Muslim, Sunan Abu Dawud, Jami' At-Tirmidhi, Sunan An-Nasa'i, Sunan Ibn Majah) contain over 30,000 individual narrations, each with a chain of transmission (isnad) and a body text (matn).

NLP annotation tasks on Hadith literature include:

• Isnad NER: Extracting the chain of narrators from the transmission chain text. This is a specialist NER task because narrator names follow classical Arabic naming conventions (kunya, laqab, nasab) that differ from modern Arabic naming and require knowledge of classical Arabic biographical dictionaries (rijal literature).
• Matn classification: Labelling the body text of each Hadith by topic, legal domain, and contextual application — tasks that require Islamic studies competency.
• Sahih/Da'if classification: While the canonical hadith grading is established, AI platforms building recommendation systems need to annotate for relative strength within the gradations, which involves hadith science (mustalah al-hadith) expertise.
• Cross-collection deduplication: Identifying narrations that appear in multiple collections as the same underlying Hadith — a semantic similarity task that requires knowledge of how the same event is described in different narrative styles across collections.

Classical Islamic jurisprudence texts (fiqh) — the Hanafi, Maliki, Shafi'i, and Hanbali madhab texts — add another annotation domain with a distinct vocabulary of legal concepts, case reasoning structures, and scholarly commentary. Fiqh NLP is nascent but growing rapidly in the GCC as Islamic fintech requires AI systems that can analyse contracts for Sharia compliance.

Case Study: GCC Islamic Knowledge Platform — Quranic Search Accuracy From 51% to 83%

A GCC-based Islamic knowledge platform serving users across Saudi Arabia, UAE, Kuwait, and Malaysia engaged AI Taggers in early 2026 to improve the quality of their Quran and Hadith search system. The platform had deployed a fine-tuned AraBERT model for semantic search but user satisfaction surveys showed significant frustration with search results — users felt the system was returning passages that were technically relevant by keyword overlap but wrong by Islamic scholarly context.

An audit of 1,000 search query-result pairs found the core issue: the existing training data had been annotated by native Arabic speakers without classical Arabic or Islamic science training. Annotators had correctly identified surface-level semantic relevance but had missed the distinction between passages that address the same topic and passages that address the same legal question — a distinction that matters enormously to users of an Islamic legal research tool.

The annotation project involved two phases:

1. Quranic thematic annotation: 6,240 verse-level annotations across 15 thematic categories (worship, ethics, legal rulings, family law, social justice, eschatology, stories of prophets, etc.) and 8 legal-domain sub-categories for verses with fiqh implications. All annotations reviewed by two scholars with ijaza (certification) in Quranic sciences.
2. Query-passage relevance annotation: 12,000 query-passage pairs annotated for three relevance levels (directly answers, thematically related, not relevant) and one specialist dimension (Islamic scholarly consensus on the connection). Annotators held degrees in Islamic studies or Sharia law.

Post-deployment results after retraining on the annotated corpus:

• Quranic search precision at rank 1: 51% → 83%
• Hadith search nDCG@5: 0.54 → 0.79
• User satisfaction (5-point scale): 2.9 → 4.3
• Query reformulation rate (proxy for search failure): 41% → 17%

The project required 22 weeks because of the scholar review layer — finding scholars with both Islamic science credentials and annotation protocol familiarity is the primary bottleneck in classical Arabic AI projects. The annotated corpus is now maintained with quarterly refresh cycles as new search query patterns emerge.

Annotator Qualification Requirements: What Actually Works

The annotator requirement for Classical Arabic NLP is not “native Arabic speaker.” It is not even “university-educated Arabic speaker.” The correct requirement is task-specific, and understanding the distinction prevents expensive annotation failures.

For morphological tagging of Classical Arabic texts: Annotators need formal training in classical Arabic grammar (nahw and sarf at the level of Al-Ajrumiyyah or higher). This is typically found in annotators with Islamic education backgrounds — not necessarily full Islamic science degrees, but exposure to classical Arabic grammar instruction. Egyptian and Saudi universities with strong Islamic studies faculties produce annotators with this background.

For NER on Hadith and historical texts: Annotators additionally need familiarity with classical Arabic naming conventions — the kunya (Abu X / Umm X form), the laqab (honorific), the nasab (genealogical chain). Without this, annotators systematically miss entity boundaries and misattribute names. A one-day calibration training on naming conventions is not sufficient — annotators need background knowledge to identify entities in context.

For Quranic thematic and legal annotation: Annotators need Islamic studies training at the level of a bachelor's degree in Sharia or equivalent. All annotations should be reviewed by at least one scholar with ijaza in the relevant Islamic science.

For Tajweed annotation for speech AI: Annotators need Tajweed certification — specifically, an ijaza in Tajweed recitation. This is a distinct qualification from Islamic science training and requires finding a different annotator pool.

Heritage Manuscripts: The Frontier of Classical Arabic NLP

Beyond the canonical Islamic texts, an enormous corpus of Classical Arabic manuscripts exists in libraries across the Arab world, Turkey, and Iran — the collections of the King Abdulaziz Public Library, the Süleymaniye Library in Istanbul, the Bodleian Library at Oxford, and hundreds of smaller archives. The King Abdulaziz Public Library alone holds over 300,000 manuscripts, and digitisation programmes funded by Saudi Vision 2030 are making these collections available.

Manuscript NLP annotation adds two challenges beyond standard Classical Arabic annotation. First, HTR (Handwritten Text Recognition) errors: Digitised manuscripts must be transcribed via HTR before NLP annotation begins. HTR models for Arabic manuscripts achieve 80–92% character accuracy but introduce systematic errors that annotators must understand to read correctly. Second, scribal variation: Pre-print manuscripts were copied by hand, and scribal traditions varied by region and period — spelling conventions, word boundaries, and abbreviations differ across the manuscript tradition.

Manuscript annotation projects in 2026 are typically structured as three-layer pipelines: HTR transcription → Classical Arabic annotator review and correction → domain-expert annotation for NLP tasks. Annotation throughput is significantly lower than for digital text — 200–400 tokens per annotator-hour for manuscript review, compared to 800–1,200 tokens per hour for standard Classical Arabic digital text.

What to Ask Before Commissioning Classical Arabic Annotation

When evaluating an Arabic NLP annotation partner for classical or Quranic Arabic projects, these questions surface genuine capability:

• “What is the educational background of your classical Arabic annotators?” If the answer is “native Arabic speakers with university education,” ask specifically about classical grammar training. A computer science graduate with native Arabic is not qualified for classical annotation.
• “Do you have annotators with Islamic science degrees?” Separate from classical grammar training — needed for theologically significant tasks.
• “Do you have access to Tajweed-certified annotators?” Needed specifically for Quran recitation AI tasks.
• “How do you handle scholar review for sensitive theological annotation?” A partner without a scholar review process should not be trusted with Quranic annotation.
• “Can you provide inter-annotator agreement statistics on a sample classical text?” Classical Arabic morphological annotation agreement below 0.75 kappa indicates annotator qualifications are insufficient.

For broader Arabic NLP dataset building, see our post on sourcing and building Arabic NLP datasets and our guide to Arabic instruction-tuning data. Our Arabic NLP annotation service covers the full spectrum from modern dialect annotation through classical and Quranic text, with qualified annotator pools for each domain.

Frequently Asked Questions