Moroccan Darija Annotation: The Hardest Arabic Variant to Get Right

Why Darija Is Not "Just Arabic"

When AI teams describe their product as "Arabic-language", they almost always mean Modern Standard Arabic plus one or two major dialects — typically Egyptian, because Egyptian Arabic is the most-understood pan-Arab variety and the most-represented in Arabic crowdsource annotation pools. Moroccan Darija is rarely included in that assumption, and it is genuinely different enough that treating it as a dialect of Arabic in the same way Egyptian is a dialect is a category error.

Linguists classify Darija as a Maghrebi Arabic variety shaped by three distinct historical layers: pre-Hilalian Arabic spoken in North Africa before the 11th-century Banu Hilal migration, the Tamazight (Berber) substrate spoken by the indigenous population, and — uniquely among Arabic dialects — centuries of French colonial influence that has made French genuinely constitutive of educated urban Moroccan speech, not a prestige overlay.

A 2023 study by researchers at Mohammed V University, Rabat, estimated that MSA-trained transformer models transfer at 34–42% accuracy on Moroccan Darija classification tasks, compared to 65–78% on Gulf Arabic tasks and 70–82% on Egyptian Arabic tasks, using the same MSA base. That gap is not a data volume problem. It reflects genuine linguistic distance — Darija has different phonology, different morphology, different core vocabulary, and different pragmatics from MSA and from other Arabic dialects.

The Four Features That Break Standard Arabic NLP on Darija

1. French lexical insertion. In educated urban Moroccan usage — Casablanca, Rabat, Marrakesh — 20–40% of vocabulary in a typical written message may be French words. These appear in three forms: Arabic-script transliteration of French words (كمبيوتر from "computer" is standard Arabic; more distinctively Darija would use French words transliterated with Moroccan phonology), Latin-script French words embedded in otherwise Arabic-script sentences, and hybrid coinages that blend Arabic morphological patterns with French roots. Models trained on Arabic-only data cannot parse any of these forms correctly.

2. Tamazight substrate vocabulary. Tamazight (Berber) is the indigenous language family of North Africa, spoken by approximately 40% of Morocco's population as a first or co-first language. Even Moroccans who don't identify as Amazigh use Tamazight-origin vocabulary in everyday speech. Common Darija words with Tamazight roots — including words for body parts, household items, and agricultural terms — have no Arabic etymology and are not found in any Arabic NLP resource.

3. Orthographic instability. Darija has no standardised written form. This is not a minor variation issue — it is a systematic feature of the language that annotation guidelines must explicitly address. The same Darija word can appear in Arabic script, Latin script, a mix of both (Arabizi), and multiple different Arabic-script spellings reflecting individual orthographic choices. A social media post in Darija may switch between scripts mid-sentence. Annotation guidelines that do not provide explicit rules for each orthographic form produce inter-annotator disagreement rates that make the resulting labels unusable.

4. Code-switching at the lexical, syntactic, and pragmatic levels. Darija code-switching is not decorative. It conveys register, social identity, and topic — switching to French signals formality, education, or professional context; switching to Classical Arabic signals religious or official register; switching to Tamazight signals ethnic solidarity. Models that cannot parse the switching signal lose significant semantic information. Annotation for sentiment, intent, or topic classification must account for the register-signalling function of language choice, not just the content of the switched items.

Who Is Building Moroccan AI in 2026?

Morocco's digital economy is growing fast. The country's Digital Morocco 2030 strategy, backed by EUR 2.25 billion in planned investment, has set targets for AI-assisted government services, digital banking access, and Arabic-language EdTech. The French-Moroccan and Spanish-Moroccan technology partnerships are producing a distinct set of AI product companies that require Darija capability.

Three sectors are driving the majority of current Darija annotation demand. Moroccan telecom operators (Maroc Telecom, Orange Maroc, Inwi) are deploying Darija-capable customer service chatbots to serve customers who prefer Darija over French or MSA in digital channels. Moroccan fintech companies — enabled by Bank Al-Maghrib's regulatory sandbox — are building credit-scoring, fraud detection, and digital wallet AI that must process Darija customer messages and social signals. And international technology companies expanding into Morocco for the first time are discovering that their French-language or MSA-language products do not serve the majority of Moroccan users adequately.

The Arabic data labeling needs of these teams are substantial and growing. By 2025, Morocco had more than 45 million internet users with 90%+ mobile penetration, generating Darija digital text at scale. The gap between that data volume and available Darija-specific annotation resources is what creates the product quality problems these teams encounter.

Case Study: Moroccan Telecom Chatbot Sentiment Model

A Moroccan telecom operator building a Darija customer service AI came to us with a three-class sentiment classifier — positive, neutral, negative — trained on a multilingual Arabic training set that included MSA, Egyptian, and Levantine data but no Darija-specific examples. The model was achieving 58.3% accuracy on Moroccan customer messages, well below the 80% threshold needed for production deployment.

The failure modes were concentrated in three areas. First, code-switched French negation was being missed: "مزيانة مش vraiment" (good, not really) was classified as positive because the model weighted مزيانة (good/nice in Darija) and did not understand that the French "pas vraiment" fragment negated it. Second, Darija-specific intensifiers drawn from Tamazight and Moroccan colloquial Arabic were being classified as neutral because they did not appear in MSA or other Arabic training data. Third, sarcasm expressed through a combination of Darija exaggeration markers and French register-switching was systematically misclassified as positive.

We ran an annotation programme in three phases. Phase one: 4,000 de-identified Moroccan customer messages annotated with three-class sentiment by native Darija-speaking annotators from Casablanca and Rabat, with guidelines explicitly addressing code-switching, Tamazight vocabulary, and sarcasm markers. Phase two: dual-QA by senior Moroccan linguists with NLP annotation experience, with adjudication on 312 disagreements concentrated in the sarcasm and negation categories. Phase three: 800-sentence adversarial set deliberately over-sampled for code-switching, negation, and sarcasm — annotated to train the model on its own systematic failure patterns.

Results after fine-tuning on the Darija-specific dataset: overall sentiment accuracy from 58.3% to 86.1%; sarcasm detection recall from 22.7% to 74.3%; code-switched negation handling from 31.4% accuracy to 83.7%; and inter-annotator agreement (Cohen's kappa) across annotators from κ=0.61 (pilot) to κ=0.84 (main production batch after guideline iteration). The operator deployed the model to a pilot of 200,000 monthly customer interactions, with manual review rate dropping from 34% to 11% of interactions within six weeks.

Annotation Task Types for Darija AI Products

Darija annotation spans different modalities depending on the product type. Social media and customer service AI primarily needs text annotation; voice AI and call centre applications need speech annotation; content moderation products may need both.

Sentiment and opinion mining. The most common entry point for Moroccan social media and customer service AI. Three-class (positive/neutral/negative) annotation is common for basic applications; more sophisticated products use fine-grained emotion labels (frustrated, delighted, confused, suspicious) or aspect-based sentiment (rating the service experience separately from the product experience). All sentiment annotation for Darija must address the code-switching and sarcasm challenge described above.

Named entity recognition. Moroccan NER requires an entity taxonomy that reflects Moroccan geography (medinas, douar names, wilaya administrative divisions), Moroccan organisations (government ministries, state companies, major corporations), and Moroccan person naming conventions (Arabic given names, Amazigh names, French names). NER models trained on MSA or pan-Arab data systematically miss Moroccan-specific entities.

Content moderation. Content moderation for Moroccan platforms requires Darija-native judgement on hate speech that references Moroccan political fault lines, ethnic communities (Amazigh, sub-Saharan African migrants), and religious discourse patterns specific to Moroccan Islamic practice. Non-Moroccan Arabic moderators consistently under-flag and over-flag categories that require cultural context — producing both moderation failures and false positives that damage platform trust.

ASR transcription for Moroccan Arabic speech. Darija speech transcription for call centre AI and voice products requires decisions about how to transcribe code-switched French — whether to transcribe in Arabic script, Latin script, or a hybrid — and how to handle Tamazight vocabulary. These decisions must be made explicitly in transcription guidelines and applied consistently to produce training data usable for ASR model fine-tuning. See our guide on multilingual speech transcription annotation for the broader framework.

Building Darija Annotation Guidelines

Darija annotation guidelines require more upfront investment than equivalent tasks in other Arabic dialects or European languages. The orthographic instability issue alone typically requires a dedicated section in the guidelines: how to handle the same word appearing in Arabic script, Latin script, and Arabizi within a single dataset; which script representation takes precedence for annotation; how to handle ambiguous transliterations.

Code-switching rules are the second major guideline investment. For each annotation task, the guidelines must specify: how to annotate items that are syntactically Darija but lexically French; how to handle items where the language boundary is mid-word (blending Arabic morphology with a French root); and what to do with Arabizi transliterations of French words that an Arabic-trained model would parse as Arabic.

Sarcasm and irony handling in Darija require specific positive examples in the guidelines. Moroccan social media sarcasm has recognisable markers — specific particles, exaggeration patterns, and code-switching signatures — that need to be explicitly enumerated with annotated examples for annotators to apply consistently.

Guideline development for a new Darija NLP task typically takes 3–4 weeks, compared to 1–2 weeks for equivalent MSA tasks and 2–3 weeks for other Arabic dialect tasks. The extended timeline reflects the need to address orthographic, code-switching, and sarcasm rules that MSA guidelines can omit. Teams that skip this investment and go directly to annotation production find that inter-annotator agreement is low (κ below 0.65), guidelines must be rewritten mid-project, and early annotation batches must be discarded — costing more time and money than the upfront guideline investment would have required.

Annotator Requirements for Darija Projects

Minimum annotator profile for Darija text annotation: native Darija speaker from Morocco (not the Moroccan diaspora, whose Darija may be archaic or region-specific in ways that do not reflect current usage); ability to read and write both Arabic and French at a functional level (to handle code-switched content); and familiarity with Moroccan social media conventions.

For higher-stakes tasks — political content moderation, financial document annotation, healthcare communication AI — domain-background annotators produce substantially higher inter-annotator agreement. A study from LREC-COLING 2024 on Maghrebi Arabic NLP found that domain-background annotators produced 19 percentage points higher Cohen's kappa on financial Darija NER versus general-population native Darija speakers.

Geographic mix within Morocco also matters for some tasks. Darija has regional variation between Casablanca-Rabat urban varieties, Marrakesh and southern varieties, and northern varieties influenced by Spanish in addition to French. For products targeting a national Moroccan audience, annotation teams should include speakers from multiple regions rather than relying exclusively on the Casablanca-Rabat variety that dominates commercial media.

For multilingual annotation programmes covering Darija alongside other language localisation tasks, it is worth noting that Darija annotators cannot be cross-trained on Gulf Arabic or Levantine tasks without significant reorientation — the dialect distance is too large for productive cross-task assignment.

Related Reading

• Arabic Sentiment Analysis: The Complete Guide for MENA AI Teams (2026)
• What Does End-to-End Arabic Data Labeling Look Like? (Project Case Study)
• Where Do Arabic NLP Datasets Come From — and How Do You Build Your Own?

Frequently Asked Questions