Iraqi Arabic Content Moderation: What Models Get Wrong Without Native Annotators

Why Iraqi Arabic Breaks Content Moderation Models

Arabic content moderation research and commercial classifier development has concentrated on the three highest-volume Arabic dialect regions: Gulf Arabic (Saudi Arabia, UAE, Kuwait), Egyptian Arabic, and Levantine Arabic (Lebanon, Syria, Jordan). Iraqi Mesopotamian Arabic — despite being spoken by approximately 40 million people — is underrepresented in Arabic hate speech, harassment, and harmful-content corpora by a factor of 8–12 compared to Egyptian Arabic and 5–7 compared to Gulf Arabic, based on corpus documentation from the OSACT (Offensive Speech in Arabic Corpus) shared tasks and the ArHateful dataset documentation (Mulki et al., 2021).

The representational gap compounds a register gap. Iraqi Arabic has developed a distinctive set of rhetorical registers since 2003 — shaped by post-occupation political discourse, sectarian dialogue, and the specific vocabulary of Iraqi civil society — that simply do not appear in Gulf or Egyptian Arabic content moderation training data. A classifier trained on those corpora encountering Iraqi Arabic text is not just operating on an unseen dialect; it is encountering a rhetorical tradition it has no coverage for at all. The practical result is false positive rates of 38–53% on Iraqi Arabic political and social commentary — content that is being incorrectly removed — alongside false negative rates of 29–44% on genuinely harmful content expressed in Iraqi-specific idiom that the classifier cannot recognise as harmful.

Research from the SemEval 2022 Arabic hate speech detection task found that models fine-tuned on multi-dialect Arabic data still produced F1 scores 23–31 points lower on Iraqi dialect hate speech than on Egyptian or Gulf Arabic hate speech under identical evaluation conditions (Mubarak et al., 2022). For platforms serving Iraqi users, these are not acceptable error rates — they produce both over-removal of legitimate political speech and under-removal of genuine harassment that reaches users.

Five Failure Modes in Iraqi Arabic Content Moderation

1. Post-2003 political hyperbole as routine commentary

Iraqi political discourse since 2003 has developed a hyperbolic register characterised by extreme condemnation, conspiratorial framing, and calls for dramatic action that are routine in Iraqi political commentary but would constitute incitement in the political discourse of most other Arabic-speaking countries. Phrases that translate literally as calling for the removal or destruction of political figures are in Iraqi political context expressions of ordinary opposition frustration, used daily by mainstream commentators without any expectation of literal interpretation. Arabic content moderation models trained on Saudi or Egyptian political content do not have exposure to this rhetorical register and classify routine Iraqi political hyperbole as incitement or calls-to-violence at rates of 45–60%.

This is not a marginal edge case — it affects the majority of Iraqi political commentary on any topic related to the political system, governance, or public figures. A platform that automatically removes content at a 45–60% false positive rate on political commentary is effectively suppressing Iraqi political discourse at scale, creating trust and retention problems with Iraqi user bases that compounds over time as users learn the platform cannot reliably evaluate Iraqi political speech.

2. Tribal-register violence as social ritual

Iraqi tribal society maintains a distinct rhetorical register where formulaic expressions of violent intent are markers of tribal loyalty, honour, and social bond rather than literal threats. Phrases from the tribal register that translate into explicit threats of physical violence or destruction are, in their social context, expressions of commitment to one's family or tribal group — the equivalent of a fraternal oath rather than a threat to an adversary. Arabic content moderation models trained on Egyptian or Gulf Arabic have no exposure to Iraqi tribal-register idiom and classify these formulaic expressions as genuine threats at rates of 51–67%, because the literal semantic content of the words aligns with threat classification categories regardless of the social register in which they are used.

Native Iraqi annotators from tribal-connected backgrounds reliably distinguish tribal-register ritual language from literal threat, drawing on contextual signals — the relationship between speaker and addressee, the presence of tribal-affiliation markers, the formulaic nature of the phrasing — that no current Arabic moderation model has been trained to detect. The same phrase appearing in a different social context, addressed to a political opponent rather than a family member, would be correctly classified as a threat by the same native annotator.

3. Mesopotamian sarcasm encoded as extreme praise

Baghdadi and broader Mesopotamian Arabic has a distinctive sarcastic register where genuine criticism, contempt, or complaint is expressed through elaborate, seemingly sincere praise — the more extreme the praise, the stronger the underlying criticism signal is to native readers. A comment describing a politician as "the greatest genius in the history of human civilisation" in Baghdadi colloquial is, in most contexts, withering sarcasm. Arabic content moderation models trained on sentiment-labelled Gulf or Egyptian Arabic data do not recognise this polarity inversion in the Mesopotamian sarcastic register and classify the text as positive sentiment, missing the harassment or coordinated-negative-campaign signal entirely.

The false negative rate on Iraqi sarcastic harassment — content that appears to be praise but functions as organised denigration — is estimated at 38–52% for commercial Arabic moderation classifiers, based on manual review of moderation decisions on Iraqi social media content (Zampieri et al., 2022, Arabic subtrack). Native Iraqi annotators reliably detect Mesopotamian sarcasm at IAA rates of 91–94%; annotators from other Arabic dialect backgrounds detect it at rates of 67–73%, a gap that is material for production moderation accuracy.

4. Kurdish code-switching and cross-language harmful content

Northern Iraqi users — in Mosul, Kirkuk, Erbil, and the diaspora from these regions — frequently code-switch between Iraqi Arabic and Kurdish within a single post or comment thread. Harmful content, harassment campaigns, and incitement appear in both the Arabic and Kurdish segments of mixed-language northern Iraqi content. Arabic-only moderation classifiers produce a 100% miss rate on harmful content in Kurdish-language segments: the model has no Kurdish training data and produces no classification output on Kurdish text, leaving the Kurdish portions of northern Iraqi content completely unevaluated. For platforms with significant northern Iraqi user bases, this represents a structural coverage gap that cannot be addressed by improving the Arabic moderation model alone.

5. Persian and Turkic loanwords carrying specific offensive register

Iraqi Arabic has a substantially larger Persian and Turkic loanword vocabulary than Gulf or Egyptian Arabic, reflecting centuries of Ottoman and Persian cultural contact. Some of these loanwords — particularly older Turkic terms — carry offensive or demeaning connotations in Iraqi colloquial usage that are entirely absent from their etymology and from their usage in any other Arabic-speaking context. Arabic NLP models trained on Gulf or Egyptian corpora have no coverage for the offensive register of these Iraqi-specific loanwords and classify text containing them as neutral or positive, producing false negatives on content that native Iraqi readers immediately recognise as derogatory or demeaning.

The Iraqi Arabic Moderation Data Gap

Building effective Iraqi Arabic content moderation annotation data requires native Iraqi annotators who can evaluate text in the full sociopolitical context of Iraqi Mesopotamian Arabic — not annotators from other Arabic dialect backgrounds applying generic Arabic moderation guidelines. The OSACT shared tasks and the Arabic hate speech detection literature consistently show that cross-dialect annotation of Arabic harmful content produces IAA scores 15–22 points below same-dialect annotation, with the accuracy gap concentrated precisely in the politically and culturally specific content categories where Iraqi Arabic diverges most sharply from MSA and other major dialects.

Our Arabic NLP annotation service provides Iraqi Arabic content moderation annotation with native Baghdadi, Basrawi, and Mosuli annotators. Annotation protocols include explicit decision trees for post-2003 political hyperbole, tribal-register language reference, Mesopotamian sarcasm indicators, and Arabic-Kurdish bilingual review routing for northern Iraqi mixed-language content.

Case Study: Baghdad News Aggregator — False Positive Rate From 46.3% to 11.8%

A Baghdad-based news aggregator and discussion platform — serving approximately 2.8 million monthly active Iraqi users across political commentary, local news discussion, and community forums — deployed a commercial Arabic content moderation classifier to automate policy enforcement on user comments and community posts. The classifier was described by its vendor as "pan-Arabic" but had been trained primarily on Egyptian and Gulf Arabic content moderation data with no Iraqi-specific training examples.

Before: The commercial classifier produced a 46.3% false positive rate on Iraqi political commentary — meaning nearly half of all Iraqi political discussion content was incorrectly flagged for removal. The false positive concentration was highest on content discussing national government performance (62.7% false positive rate), electoral politics (58.1%), and sectarian political parties (71.4%). On genuinely harmful content — harassment campaigns targeting journalists, coordinated denigration of minority community members, incitement related to sectarian tension — the classifier produced a 37.2% false negative rate, missing more than a third of harmful content in the moderation queue.

User appeal volume was 4,100 appeals per week, with 91.3% of successfully appealed removals overturned — confirming that the automatic removal decisions were incorrect at rates far above the platform's acceptable threshold. The moderation team was spending 34% of total manual review time processing appeals for incorrect automatic removals rather than evaluating new content, creating a backlog in which genuinely harmful content waited 8–14 hours for human review.

The moderation annotation project delivered 94,000 labelled Iraqi Arabic items across six content categories: political commentary, civic discussion, community interaction, cultural content, commercial content, and cross-dialect northern Iraqi content. Annotation was conducted by a team of fourteen native Iraqi annotators — nine Baghdadi-native, three Basrawi-native, and two Mosuli Arabic-Kurdish bilingual — following protocols with explicit decision trees for post-2003 political hyperbole, tribal-register expression reference, Mesopotamian sarcasm coding rules, and Kurdish-segment routing. Inter-annotator agreement on the final annotated dataset reached 91.4% kappa across all six content categories.

After fine-tuning on the Iraqi moderation dataset: The false positive rate on Iraqi political commentary fell from 46.3% to 11.8% — within the platform's acceptable threshold for automatic enforcement. False negative rate on genuinely harmful content fell from 37.2% to 9.4%. Appeal volume dropped from 4,100 to 870 per week — an 78.8% reduction. The moderation team's time spent processing appeals fell from 34% to 8% of total manual review hours, redirecting that capacity to human review of genuinely ambiguous content.

Mean time to action on genuinely harmful content fell from 8–14 hours to 2.3 hours, driven by the reduction in false-positive noise in the moderation queue. User trust metrics — measured via post-session survey — improved 22.4 percentage points on the statement "this platform allows me to express my political views without unfair removal." The annotation project cost AUD $38,000 for annotation, protocol development, quality assurance at 8% blind-sample review rate, and delivery in the classifier fine-tuning format required by the platform's moderation infrastructure.

What Iraqi Arabic Content Moderation Annotation Requires

Effective Iraqi Arabic moderation annotation requires protocol elements that generic Arabic moderation guidelines do not include.

Post-2003 political hyperbole decision tree. Annotation guidelines must include explicit decision criteria distinguishing routine Iraqi political exaggeration from genuine incitement — with reference examples drawn from Baghdadi and broader Iraqi political discourse at each decision node. Without this, native annotators apply inconsistent standards on the most common false-positive category in Iraqi content moderation, producing IAA scores of 71–78% on political commentary rather than the 89–93% achievable with explicit guidelines.

Tribal-register lexicon reference. Annotation guidelines must provide a reference list of common tribal-register phrases and their social meaning, with contextual criteria specifying when tribal-register expression should be classified as benign (loyalty expression, social ritual) versus when the same phrase in a different relational context should be classified as a genuine threat. This lexicon cannot be derived from generic Arabic resources — it requires input from Iraqi annotators with tribal-context knowledge.

Mesopotamian sarcasm coding rules. Guidelines must include explicit rules for detecting the Mesopotamian sarcastic-praise register: indicators of ironic polarity inversion (extreme praise intensity as a negative signal), contextual markers of sarcasm (account history, comment positioning, subject matter), and minimum confidence threshold for sarcasm classification. Sarcasm coding rules specific to Baghdadi colloquial are required — Levantine or Egyptian sarcasm detection heuristics do not transfer reliably.

Kurdish-language segment routing. Annotation workflows for northern Iraqi content must specify that any Kurdish-language segment in a post or comment must be routed to an Arabic-Kurdish bilingual annotator before a final moderation classification can be issued for the item. Annotation of the Arabic-language segments by Arabic-only annotators followed by a note of "Kurdish segment unevaluated" is not an acceptable annotation output for platforms where harmful content regularly appears in the Kurdish-language portions of northern Iraqi mixed-language posts.

Our Arabic NLP annotation service provides all four protocol elements as standard for Iraqi Arabic content moderation projects, with annotation output validated against a 8–10% blind-sample quality assurance process and delivered in the label format required by your moderation classifier fine-tuning pipeline.

Related Reading

• Iraqi Arabic Sentiment Analysis: What Models Get Wrong Without Native Annotators
• Iraqi Arabic Chatbot Intent Annotation: What Models Get Wrong Without Native Annotators
• Gulf (Khaleeji) Arabic Content Moderation: What Models Get Wrong Without Native Annotators
• Where Do Arabic NLP Datasets Come From — and How Do You Build Your Own?
• Arabic Data Labeling Service

Frequently Asked Questions