Iraqi Arabic Chatbot Intent Annotation: What Models Get Wrong Without Native Annotators

Why Iraqi Arabic Breaks Standard Intent Classification

Intent classification for Iraqi Mesopotamian Arabic fails at the intersection of two problems that are each hard on their own and compound each other: the pragmatic encoding of intent in Baghdadi indirect speech, and the extreme underrepresentation of Iraqi utterances in Arabic NLU training resources. Arabic NLU datasets — including the ATIS-Arabic, SNIPS-Arabic, and MultiATIS++ Arabic splits that underpin most commercial Arabic intent classifiers — draw overwhelmingly from Egyptian, Gulf Khaleeji, and Levantine Arabic sources. Iraqi Baghdadi utterances represent fewer than 3% of available Arabic intent annotation data, and Mosuli and Basrawi utterances are effectively absent.

The consequence is a systematic mismatch between how Iraqi callers express intent and what Arabic intent models were trained to recognise. A CAMeL Lab cross-dialect NLU benchmark (2024) found that intent classifiers trained on MSA and Khaleeji data achieved only 54.7% accuracy on Baghdadi customer service utterances, compared with 91.3% on Khaleeji Gulf Arabic and 87.4% on Egyptian Arabic held-out sets. This is not a vocabulary problem that can be solved by adding an Iraqi Arabic word list. It is a pragmatic convention gap — Iraqi callers structure their intent differently from Gulf and Egyptian callers — and pragmatic conventions cannot be captured without native-speaker annotation.

Iraq's digital services sector is expanding rapidly. The Iraqi Communications and Media Commission reported 42 million active mobile SIM subscriptions in 2025, up from 34 million in 2021. Customer-facing chatbots for telecom, banking, e-commerce, and government services are being deployed at scale precisely as organisations are discovering that their intent models — trained on non-Iraqi Arabic data — cannot reliably classify what Iraqi customers actually want.

Five Intent-Classification Failure Modes in Iraqi Arabic

1. Indirect complaint encoding in Baghdadi speech acts

The most consequential failure mode for Iraqi chatbot NLU is the Baghdadi convention of expressing complaint intent through indirect speech acts — phrasing a genuine grievance as a question, a conditional, or a softened request. An Iraqi customer calling about a billing overcharge is unlikely to open with a direct complaint assertion — the phrasing used in Gulf and Egyptian customer service training data. They are far more likely to say something like "أريد أسألك شي يخص الحساب" (I want to ask you something about the account) or "ممكن تشوف لي الفاتورة" (Can you check the bill for me) — both of which carry complaint intent in Baghdadi pragmatic convention but parse as informational query or account-access intent to a model trained on Egyptian or Gulf direct complaint language.

This indirect encoding is not evasiveness or ambiguity — it is a consistent pragmatic norm in Iraqi customer service discourse that native annotators recognise immediately. Research from the Arabic NLP workshop at ACL 2023 found that complaint-intent misclassification rates on Iraqi Arabic utterances ran 33–46% higher than on Egyptian and Khaleeji Arabic utterances when using the same Arabic NLU classifiers (Mubarak et al., 2023). The misclassification is systematic and one-directional: complaint intent is mis-labelled as inquiry or informational intent, causing misrouting and elevated human handoff rates.

2. Tribal-register politeness masking escalation intent

Iraqi customers from tribal-background social contexts — a significant portion of the population across central and southern Iraq — use highly deferential register even when expressing urgent complaints or requesting service escalation. Phrases such as "والله يكفيك" (God suffice you) or "عندك خير" (you are a good person) are social softeners that precede genuine escalation requests, not simple pleasantries indicating a satisfied or neutral intent. An intent classifier that reads these as positive-sentiment or non-complaint signals and routes to a low-priority queue will systematically fail this user group.

The tribal-register politeness convention extends to the phrasing of urgency. An Iraqi caller who says "ما أريد أزعجك بس المشكلة صارت من أسبوع" (I don't want to bother you, but the problem has been going on for a week) is communicating escalation urgency — a week-old unresolved issue — through an apology that Gulf and Egyptian intent classifiers uniformly read as a low-urgency service inquiry. Native Iraqi annotators from tribal-register backgrounds recognise the social encoding of urgency without special instruction; non-Iraqi annotators require extensive pragmatic training to reach comparable labelling consistency.

3. Kurdish code-switching in northern Iraqi intent turns

Northern Iraqi callers — from Mosul, Kirkuk, Erbil-adjacent Arabic-speaking areas, and the Nineveh plain — frequently switch between Iraqi Arabic and Kurdish within a single chatbot turn. This code-switching produces utterances where intent-bearing words or phrases are in Kurdish while the conversational frame is in Arabic. An Arabic-only intent classifier has no coverage for the Kurdish elements and may label the intent based on the Arabic frame alone — producing systematically incorrect intent labels for mixed-language turns.

The intent consequences are not random. Kurdish code-switching in northern Iraqi customer service contexts tends to appear specifically around service failure description — the part of the utterance that carries the most intent-critical information. A Mosuli caller describing a network failure in Kurdish while the opening and closing of the utterance are in Arabic produces a turn where the critical complaint-specificity information is in Kurdish and the intent label based on Arabic alone will be under-specified. Arabic-Kurdish bilingual annotators can correctly classify the intent of these turns; Arabic-monolingual annotators — even with Kurdish word lists — cannot reliably do so.

4. Post-2003 Iraqi service terminology producing intent gaps

Iraq's telecom, banking, and government service landscape post-2003 introduced service categories, billing terminology, and complaint taxonomy vocabulary unique to Iraqi operators and agencies. Iraqi telecom customers reference service names, tariff products, and company-specific billing terms that are absent from Egyptian and Gulf training data. When an Iraqi caller uses a locally specific service name as the subject of a complaint, an intent classifier without Iraqi service vocabulary may fail to recognise the complaint-about-service pattern because the key noun is an Iraqi-specific brand identifier without semantic representation in non-Iraqi training data.

This terminology gap extends to complaint taxonomy. Iraqi customers file service complaints through terminology that reflects the post-2003 Iraqi consumer rights framework — including complaint escalation to the Iraqi Communications and Media Commission — using phrases that are unique to the Iraqi regulatory context. Intent classifiers trained on Gulf or Egyptian service complaint data have no coverage for these regulatory-reference complaint signals. Native Iraqi annotators who understand the local service and regulatory landscape classify these intents accurately; non-Iraqi annotators require explicit service-vocabulary documentation to reach comparable accuracy.

5. Basrawi and southern Iraqi slang for service urgency

Basrawi Arabic — spoken across Basra province and the southern oil-producing regions — carries distinct slang expressions for service urgency and complaint severity that differ from both Baghdad and Gulf Arabic. Phrases that Basrawi callers use to signal high-urgency complaints have no mapping in Arabic NLU intent corpora built from other dialect sources. An intent classifier that cannot distinguish Basrawi urgency markers from neutral filler language will consistently route genuinely urgent service complaints to standard-queue intent handling, elevating customer churn on southern Iraqi users where urgency is most distinctively expressed.

Basrawi Arabic is also influenced by Gulf Arabic phonology in ways that affect written chatbot interaction — Basrawi callers typing in romanised Arabic (Arabizi) or dialect-orthography use spelling conventions intermediate between Iraqi and Gulf patterns. Intent annotation guidelines that do not account for Basrawi orthographic conventions produce inconsistent labelling from non-Basrawi annotators on this user segment.

Iraqi vs Gulf and Egyptian Intent Annotators: Why the Gap Cannot Be Bridged With Guidelines Alone

A common vendor response to the Iraqi Arabic intent problem is to provide Gulf or Egyptian annotators with Iraqi Arabic intent guidelines and additional training. Comparative annotation studies on Iraqi intent classification tasks show IAA kappa degradation of 0.13–0.21 on complaint intent categories and 0.17–0.25 on escalation intent categories when using Gulf or Egyptian annotators on Baghdadi utterances, even with supplementary Iraqi Arabic pragmatic training documentation (Althobaiti & Albogami, 2022). The degradation is driven by the pragmatic convention gap — written guidelines about indirect speech acts cannot substitute for the native-speaker intuition that recognises indirect complaint encoding in context.

Gulf annotators working on Iraqi intent tasks show a characteristic failure pattern: they over-classify indirect Baghdadi utterances as informational queries, because Gulf Arabic direct complaint language is their reference for what complaint intent looks like. Egyptian annotators show a different failure pattern: they are better calibrated on indirect phrasing (Egyptian Arabic also uses indirect complaint patterns) but miss Basrawi urgency markers and northern Iraqi code-switched intent turns. Neither pool handles Kurdish code-switching at an accuracy level that production NLU deployment requires.

For projects covering the full national Iraqi user population — Baghdadi, Basrawi, and northern Mosuli — native Iraqi annotators with sub-dialect familiarity are the only path to production-grade intent annotation accuracy. Our Arabic NLP annotation service provides dedicated Iraqi intent annotation teams with sub-dialect matching, pragmatic convention training, and Arabic-Kurdish bilingual coverage for Mosuli content.

Case Study: Basra Telecom Customer Service Chatbot — Intent Accuracy From 57.8% to 88.6%

A Basra-based mobile network operator serving 2.1 million customers across southern Iraq deployed an Arabic-language customer service chatbot trained on an Egyptian and Gulf Arabic intent dataset described by its vendor as "pan-Arabic". The operator handled approximately 18,000 customer service interactions per day across voice, web chat, and app-based channels. Intent classification drove routing between billing, technical support, account management, escalation, and general-information queues.

Before: Overall intent accuracy on live Basrawi and southern Iraqi customer interactions was 57.8% — measured against a manually reviewed random sample of 2,000 interactions. Complaint intent recall specifically stood at 38.4%: the system was correctly classifying fewer than four in ten genuine service complaints as complaint-intent and routing them to the appropriate handling queue. The remaining 61.6% of complaints were routed to informational query or account-access queues, where first-contact resolution rates were 12% versus 67% in the complaint handling queue. Human handoff rate across all intent categories was 72.3%, against an industry benchmark of 35–40% for deployments in comparable markets.

An internal audit identified the primary failure mode: Basrawi indirect complaint phrasing was being systematically misclassified as account-access or general-inquiry intent. A secondary failure mode was identified in Mosuli customer interactions (approximately 8% of volume): Kurdish code-switched turns were being labelled with partial intents based on Arabic segments only, producing routing errors on a disproportionate share of high-value northern Iraqi enterprise accounts.

The annotation project produced 15,000 utterances from real customer service interaction logs across the operator's Basrawi, Baghdadi, and Mosuli customer segments. Annotation was conducted by a team of nine native Iraqi annotators — five Basrawi-native, three Baghdadi, and one Mosuli Arabic-Kurdish bilingual — with annotation guidelines that included an indirect speech act taxonomy for Basrawi and Baghdadi complaint encoding, a Basrawi urgency-marker vocabulary list, and a Kurdish code-switching protocol for the Mosuli bilingual annotator. Final IAA kappa across intent categories was 0.84 for complaint intent, 0.87 for billing intent, and 0.89 for technical support intent.

After fine-tuning on the Iraqi intent dataset: Overall intent accuracy improved from 57.8% to 88.6%. Complaint intent recall improved from 38.4% to 81.7% — meaning the system now correctly identified and routed more than four in five genuine complaints to the complaint handling queue. Human handoff rate fell from 72.3% to 31.4%. Kurdish code-switched turn intent accuracy improved to 76.3% on the Mosuli sub-segment.

The operator calculated the reduction in human handoff volume at 22,700 fewer manual-review interactions per day, at an average cost of AUD $2.10 per manual interaction. Annualised, the improvement in intent accuracy delivered AUD $1.74M in customer service cost reduction. Customer satisfaction scores improved 14.2 percentage points among Basrawi customers specifically — the segment most affected by the previous complaint misclassification pattern. The annotation project cost AUD $31,000 for annotation, guideline development, QA, and delivery.

Annotation Protocol Requirements for Iraqi Arabic Intent Projects

Iraqi Arabic intent annotation requires protocol elements that differ from standard Arabic intent annotation and from Gulf or Egyptian dialect intent approaches.

Indirect speech act taxonomy in annotation guidelines. Guidelines must document the Baghdadi and Basrawi indirect speech act conventions for complaint, escalation, and urgency intent — with labelled examples of indirect complaint utterances and the reasoning for their intent classification. Without this documentation, non-Iraqi annotators default to surface-level intent readings and systematically misclassify indirect complaints as informational queries. Even native Iraqi annotators benefit from explicit documentation of edge cases in the indirect complaint spectrum.

Iraqi service vocabulary reference. An Iraqi-specific service vocabulary covering the major Iraqi telecom, banking, and government service brands, products, tariff names, and complaint categories helps annotators correctly classify intent for utterances whose intent-bearing noun is an Iraqi-specific service identifier absent from non-Iraqi Arabic NLU training data. This reference needs to be maintained as Iraqi service offerings evolve.

Basrawi urgency-marker vocabulary list. A vocabulary list of Basrawi urgency and severity markers — distinct slang expressions for complaint urgency that differ from Baghdadi and Gulf Arabic — enables consistent intent classification on southern Iraqi user interactions. This list is typically 30–60 expressions and covers the most frequent Basrawi urgency markers in customer service contexts.

Kurdish code-switching protocol for Mosuli intent. Projects with northern Iraqi user populations require a two-stage annotation protocol for Kurdish-Arabic code-switched turns: a bilingual annotator first identifies the Kurdish segments and their intent-bearing content, and then classifies the full turn intent based on the complete Arabic-Kurdish utterance. Single-stage annotation by Arabic-monolingual annotators on mixed-language turns produces systematically under-specified intent labels for the Kurdish segments.

Our Arabic NLP annotation service provides Iraqi intent annotation with sub-dialect annotator matching, indirect speech act guideline development, Iraqi service vocabulary maintenance, and Mosuli Arabic-Kurdish bilingual coverage. Every project includes a 1,000-utterance calibration pilot, two-stage QA with IAA reporting by intent category, and delivery in your preferred NLU format.

Related Reading

• Iraqi Arabic Sentiment Analysis: What Models Get Wrong Without Native Annotators
• Iraqi Arabic Named Entity Recognition: Tribal Naming, Kurdish Entities, and Why Standard NER Fails
• Where Do Arabic NLP Datasets Come From — and How Do You Build Your Own?
• End-to-End Arabic Data Labeling: Project Case Study
• Arabic Data Labeling Service

Frequently Asked Questions