Arabizi and Code-Switching: Annotating the Way Arabs Actually Type Online

What Arabizi Is and Why Arabic NLP Models Cannot See It

Arabizi — also written “Arabish” or “Franco-Arabic” — is Arabic written in the Latin alphabet, with numerals substituting for Arabic phonemes that lack a Latin equivalent. The letter '3' represents the pharyngeal fricative ع; '7' represents the voiceless pharyngeal fricative ح; '2' represents the glottal stop ء. It emerged in the 1990s when SMS and early messaging systems could not display Arabic script, and it remains the dominant informal written register for younger Arab users across Gulf, Egyptian, and Levantine contexts.

For an Arabic NLP model trained on Arabic-script input, an Arabizi sentence is noise. The model does not receive a degraded signal — it receives no signal at all. A sentiment classifier that correctly identifies anger in "أنا زعلان من الخدمة" will produce a random output on "ana za3lan mn el khidma" because the tokeniser has never encountered those character sequences in its training vocabulary.

According to research published by the Arabic NLP community through the ACL Anthology, Arabizi constitutes an estimated 30–45% of Arabic-language social media content in Gulf markets and up to 60% in Egyptian youth demographics. A MENA-focused platform that deploys a standard Arabic NLP stack is operating with a model that is effectively blind to the majority of its users' informal communication.

The annotation challenge is not translating Arabizi back to Arabic script — machine transliteration tools exist and handle the straightforward cases. The challenge is that Arabizi spelling is not standardised. “7abibi” and “7bibi” and “habibi” all represent the same word. An annotator unfamiliar with Arabizi conventions makes systematic errors. An annotator who does not write Arabizi natively misses pragmatic cues embedded in spelling choices.

Code-Switching: Arabic–English in a Single Sentence

Code-switching is the practice of alternating between two or more languages within a single conversational exchange or sentence. In Arabic digital communication, code-switching between Arabic (in any variety or script) and English is extremely common. It is not a sign of limited language proficiency — it is a sociolinguistic choice that signals identity, register, and in-group membership.

A Gulf user writing a negative product review might produce: “el delivery was super late, ma sar shay w el customer service useless.” This sentence contains Khaleeji Arabic (“ma sar shay” — “nothing happened”), English adjectives (“super late”, “useless”), and an English noun phrase (“customer service”). A sentiment model trained on either monolingual English or monolingual Arabic will fail to correctly aggregate sentiment across the code-switched utterance.

North African users add a third language: French. Moroccan Darija code-switching commonly involves Arabic script, Latin Darija, and French — sometimes in the same sentence. An Egyptian code-switcher uses English technical vocabulary but Arabic-script emotional expression.

According to the CAMeL Lab at NYU Abu Dhabi, code-switched Arabic–English content represents over 20% of Arabic Twitter posts and a higher proportion of Instagram and TikTok comment data from the 18–34 demographic. Any Arabic NLP product targeting that demographic — a social platform, a customer service chatbot, a content moderation system — is handling code-switched input at scale.

Where Arabizi and Code-Switching Annotation Breaks Existing Pipelines

The failure modes are task-specific, and understanding them is the starting point for building the right annotation approach.

Sentiment analysis: The most common failure is sentiment polarity reversal. Arabizi frequently uses exaggerated positive sentiment vocabulary to express sarcasm — a Gulf practice of calling something “zain zain” (good good) with particular Arabizi spelling conventions signals sarcasm, not genuine positivity. A model without Arabizi training data reads the surface positive label and inverts the correct prediction.

Content moderation: Arabizi is actively used to evade script-based content classifiers. A harmful phrase written in Arabic script triggers the classifier. Written in Arabizi with deliberate misspellings, it passes undetected. This is documented platform abuse behaviour in MENA markets. Annotating Arabizi for toxicity and harmful content is a safety-critical task.

Intent classification: Chatbot intent classifiers trained on formal Arabic MSA fail when users type naturally. A user who wants to cancel a subscription might type “abi a3mal cancel lel subscription mal3oona di” (a mix of Gulf Arabic, English, and Egyptian colloquial). Neither an Arabic intent model nor an English intent model classifies this correctly.

Named entity recognition: Person names, company names, and place names appear in Arabizi with no standardised transliteration. “Mohammed”, “Mohamed”, “Mohammad”, “M7amed” all refer to the same name. An NER model trained on Arabic-script named entities cannot link these variants.

Case Study: MENA Social Platform — From 41% to 79% Arabizi Sentiment Accuracy

A MENA-focused social content platform serving Gulf and Egyptian users engaged AI Taggers in late 2025 after their sentiment analysis pipeline produced results that were visibly wrong to human reviewers. Posts that users flagged as negative were being classified as neutral or positive. Moderation queues were missing clear violations. Customer support intent routing was misclassifying over half of informal user messages.

An audit of 5,000 user posts found that 48% were Arabizi or code-switched, with a near-equal split between Gulf Arabizi and Egyptian Arabizi. The existing model — a fine-tuned multilingual BERT variant — had been trained on Arabic-script social media data with no Arabizi coverage and no code-switching examples.

The annotation project involved three phases: script routing development (2,000 examples annotated for script type to train a routing classifier, reaching 91% accuracy after two calibration rounds); sentiment annotation (15,000 Arabizi and code-switched posts annotated for three-class sentiment and sarcasm flag, with within-dialect Cohen's kappa reaching 0.73); and moderation annotation (8,000 posts across four violation categories with attention to evasion tactics).

Post-deployment results after retraining:

• Arabizi sentiment accuracy: 41% → 79%
• Code-switched sentiment accuracy: 38% → 74%
• Moderation false negative rate on Arabizi violations: 61% → 19%
• Intent classification accuracy on informal messages: 44% → 71%

The project took 14 weeks from audit to deployment. The annotation corpus — 25,000 items — is maintained with a monthly refresh cycle as new Arabizi vocabulary and code-switching patterns emerge.

How to Build Arabizi Annotation Guidelines That Actually Work

The most common failure mode in Arabizi annotation projects is starting with annotation before establishing spelling normalisation. Arabizi has no governing body and no dictionary. Without a normalisation convention, two native Arabizi writers annotating the same dataset will produce systematically lower agreement because they are reading the same tokens differently.

Before annotation begins, guidelines for Arabizi projects must address: numeral substitution mapping (canonical mapping for 3=ع, 7=ح, etc. and how to handle ambiguous cases like 5 for خ vs. kh); vowel elision conventions (Gulf Arabizi commonly elides short vowels while Egyptian Arabizi retains them more often); code-switch boundary identification (when an English word counts as code-switching vs. an absorbed loan word); and sarcasm/irony markers specific to each dialect.

Calibration rounds — annotating 100–200 shared examples before live annotation begins — should continue until within-dialect kappa reaches at least 0.70. Below that threshold, inter-annotator variance is introducing more noise than the annotation is removing.

The Dialect Routing Layer: Why Gulf and Egyptian Arabizi Need Separate Annotators

Gulf Arabizi and Egyptian Arabizi are different enough in vocabulary, spelling conventions, and pragmatic norms that cross-dialect annotation introduces systematic errors. A native Gulf Arabizi writer reading Egyptian Arabizi will correctly identify most words but miss culturally-specific sarcasm markers, will apply Gulf spelling conventions to Egyptian text, and will misread idioms that carry different sentiment weight in each dialect.

This has a practical consequence for annotation project design: dialect routing is not optional. Before Arabizi annotation begins, a dialect identification step routes each item to the correct dialect pool. This routing step itself requires annotation — an initial batch of dialect-labelled Arabizi examples to train the routing classifier.

For projects covering both Gulf and Egyptian Arabizi, budget for two distinct annotator pools, two sets of calibration rounds, and two sets of label agreement statistics. Cross-dialect agreement numbers are not a meaningful QA signal for within-dialect annotation tasks. Levantine Arabizi and Moroccan Darija in Latin script each add a further distinct system.

Arabizi for Content Moderation: A Safety-Critical Annotation Task

Content moderation in Arabizi is not just a quality problem — it is a platform safety problem. Arabizi is actively used by bad actors to evade Arabic-script classifiers. A hate speech phrase written in Arabic script triggers an existing classifier. Written in Arabizi with deliberate numeral substitutions and vowel dropping, it passes undetected.

According to a 2024 survey of MENA platform trust and safety teams by the Global Network Initiative, Arabizi content accounts for a disproportionate share of policy violations that evade automated detection — estimated at 35–55% of false negatives in Arabic-language moderation queues for platforms without Arabizi-specific models.

Arabizi moderation annotation requires: native Arabizi fluency to read violations correctly; cultural competency for context-dependent violations; awareness of evasion tactics including deliberate misspellings and numeral substitutions; and welfare protocols equivalent to those for English-language moderation teams. It should not be treated as a variant of standard Arabic annotation.

What to Ask an Arabic NLP Annotation Partner About Arabizi Coverage

When evaluating an Arabic NLP annotation partner for Arabizi projects, these questions surface capability gaps: “What proportion of your annotators write Arabizi natively?” — a vendor that answers “our annotators are native Arabic speakers” has not answered. “How do you route Gulf vs Egyptian Arabizi?” — systematic routing, not annotator self-selection. “What do your guidelines include for spelling normalisation?” — a vendor with no normalisation convention will produce low agreement. “Can you provide within-dialect kappa statistics for Arabizi sentiment tasks?” — below 0.65, annotation quality is unlikely to produce signal above the noise floor.

For building comprehensive Arabic NLP training data, see our post on Khaleeji Arabic sentiment annotation and Egyptian Arabic chatbot annotation. Our Arabic NLP annotation service handles Arabizi, code-switched text, and all Arabic dialect varieties with the correct native-speaker annotator pools.

Frequently Asked Questions