Iraqi Arabic Sentiment Analysis: What Models Get Wrong Without Native Annotators

What Makes Iraqi Arabic a Distinct Sentiment Problem

Iraqi Mesopotamian Arabic is the dialect spoken across Iraq — from Baghdad in the centre to Basra in the south and Mosul in the north. It is structurally distinct from Gulf, Egyptian, and Levantine Arabic in ways that matter directly for sentiment analysis. Where Gulf Arabic has heavy English code-switching and Egyptian has directness shaped by Cairo media culture, Iraqi Arabic carries the linguistic legacy of Ottoman administration (Turkic loanwords), Persian proximity (Persian loanwords especially in Shia religious register), and for northern Iraq, sustained Arabic-Kurdish bilingualism.

The foundational problem is dataset underrepresentation. Analysis of seven major Arabic NLP corpora — LABR, ASTD, HARD, ArSAS, SemEval-2017 Arabic, QADI, and MADAR — finds fewer than 3% of examples from Iraqi Mesopotamian sources (Bouamor et al., 2019; Keleg & Magdy, 2023). Iraqi dialect is not merely absent from MSA benchmarks; it is largely absent from the dialect-inclusive resources that supposedly improve on MSA-only training. The result is that sentiment models described as covering “Arabic” or even “multi-dialect Arabic” perform near chance on Iraqi dialect test sets for the hardest sentiment categories.

Research from OSACT (Opinion Sentiment Analysis Shared Task for Arabic) and WANLP documents 28–44% accuracy degradation on Iraqi dialect sentiment for models trained on standard Arabic resources, with the worst failure on negative and sarcastic categories (Farha & Magdy, 2020; Abdul-Mageed et al., 2023). The Iraqi market context amplifies this: Iraq's e-commerce sector grew 34% year-on-year in 2025 according to the World Bank Iraq Digital Economy Report, creating large volumes of Iraqi dialect review and customer feedback text that existing Arabic AI cannot accurately interpret.

Five Iraqi Sentiment Patterns That Break Standard Arabic Models

1. Turkic and Persian loanwords as quality-sentiment carriers

Iraqi Arabic inherited a substantial vocabulary layer from centuries of Ottoman administration and Persian cultural proximity. Many of the quality-judgement terms used in everyday Iraqi commercial speech trace to Turkic or Persian roots that do not appear in MSA training corpora. “هواية” (very much/a lot — from Turkic “havayı”) functions as a sentiment intensifier in Iraqi. “بشلق” (cheap/worthless — from Turkic “beşlik”) carries strong negative quality evaluation. “چلو” (well/fine — Persian origin) used sarcastically signals strong negative sentiment.

These terms are invisible to any sentiment model trained on MSA or on Arabic dialects without Iraqi representation. A product review using “هواية رديء” (very bad, with Turkic-origin intensifier) registers as negative in Iraqi context but may be mis-tokenised or under-weighted by a standard Arabic sentiment system that has never seen the intensifier pattern. The problem is not simply vocabulary gap — it is that the sentiment-carrying terms require Iraqi dialect context to parse correctly.

2. Tribal register and indirect complaint expression

Iraqi social register carries strong tribal identity markers that shape how sentiment — particularly negative sentiment — is expressed in public commercial contexts. Direct complaint is frequently framed as a concern for the business's honour rather than an explicit criticism: “ما يستاهل السمعة الحلوة” (“it does not deserve the good reputation”) is a strongly negative statement in Iraqi tribal register that reads as mild or even neutral to MSA classifiers and to non-Iraqi Gulf annotators.

Iraqi tribal complaint avoids explicit criticism of individuals in public-facing text, instead routing critique through honour, reputation, and community standing. “تعبنا” (we became tired/exhausted — used to express deep dissatisfaction without naming the cause) is a Mesopotamian complaint marker that requires cultural context to score correctly. The IAA degradation from using Gulf or Egyptian annotators on Iraqi complaint text has been measured at 0.09–0.16 kappa points on negative sentiment categories in comparative annotation studies (Althobaiti & Albogami, 2022).

3. Kurdish code-switching in northern Iraqi text

Northern Iraq — Mosul, Kirkuk, Erbil — has sustained Arabic-Kurdish bilingualism. Urban commercial text from these regions shows consistent Arabic-Kurdish code-switching, with Kurdish sentiment terms embedded in Arabic sentence frames and vice versa. A product review from Mosul might express a negative quality judgement with a Kurdish intensifier in an otherwise Arabic sentence, or frame a complaint entirely in Sorani Kurdish within an Arabic-language platform thread.

No standard Arabic sentiment model handles Kurdish code-switching. Arabic-only models either fail to tokenise the Kurdish elements or route them to error categories. For platforms with northern Iraqi user bases — or for national Iraqi platforms where Mosuli and Kirkuki reviews are mixed with Baghdadi and Basrawi text — the code-switching problem is not marginal; it affects a material portion of the negative sentiment data that matters most for product and service quality monitoring.

4. Baghdadi urban sarcasm through excessive formality

Baghdadi Arabic has a distinctive sarcasm construction in urban commercial contexts: excessively formal MSA-register language used to frame a mundane complaint. A customer who writes a glowing review in formal classical Arabic for a product that clearly failed them is expressing Baghdadi sarcasm through register incongruity — the formal register applied to an everyday negative experience signals irony that is immediately apparent to a Baghdad reader and invisible to any MSA sentiment classifier that treats formal Arabic as inherently positive or neutral.

This pattern is the opposite of the sarcasm constructions documented in Khaleeji or Egyptian Arabic sentiment research, where sarcasm often uses colloquial register applied to elevated topics. Iraqi dialect sarcasm through formal incongruity requires annotators who are native Baghdadi speakers and who are explicitly trained to flag register-sentiment mismatch as a sarcasm indicator — not simply to label the surface sentiment of the words used.

5. Shia religious idiom as sentiment amplifier in southern Iraq

Southern Iraqi Arabic — particularly Basrawi — uses Shia Islamic religious idiom as a sentiment amplifier in ways distinct from the Sunni Gulf religious idiom documented in Khaleeji Arabic NLP research. Religious phrases invoking imams or holy sites function as intensifiers of both positive and negative sentiment in Basrawi register. A complaint prefaced with a phrase invoking divine justice is expressing strong negative sentiment in southern Iraqi context that a non-Shia-context annotator is likely to classify as a religious expression rather than an emotional amplifier.

This distinction means that a single “Iraqi annotator pool” that mixes Baghdadi, Basrawi, and Mosuli annotators without sub-dialect routing will produce inconsistent labels on the religious-idiom categories — not because the annotators disagree about the words, but because the sentiment weight of specific phrases varies by sub-dialect context. Annotation guidelines for Iraqi projects must include sub-dialect-specific idiom tables, not a single pan-Iraqi idiom reference.

Baghdadi, Basrawi, Mosuli: Why One Iraqi Annotator Pool Is Not Enough

Many annotation vendors treat “Iraqi Arabic” as a single category and staff projects from a general pool of Iraqi annotators. For most production use cases, this is a material quality risk. The three dominant Iraqi sub-dialects — Baghdadi, Basrawi, and Mosuli — differ in sentiment expression in ways that affect annotation accuracy on the categories that matter most.

Baghdadi is the prestige urban register, relatively more direct in commercial contexts than rural Iraqi dialects and shaped by Baghdad's role as a commercial and cultural centre. Basrawi (Southern Iraqi) is more influenced by Gulf Arabic contact and carries stronger Shia religious idiom. Mosuli has Kurdish code-switching and is influenced by northern Iraqi tribal and linguistic contact patterns. Using Baghdadi annotators for Basrawi text produces systematic under-scoring of Shia idiom sentiment weight. Using Basrawi annotators for Mosuli text produces missed Kurdish code-switched sentiment.

For national platforms serving all of Iraq, the practical protocol is a Baghdadi-dominant annotator pool with supplementary Basrawi and Mosuli annotators, plus explicit training on sub-dialect idiom tables for each pool. For regional platforms — a Basra logistics app, a Mosul education platform — the lead annotator pool should match the dominant user sub-dialect. Iraq's digital economy is growing fastest in Baghdad and Basra, with the World Bank 2025 report projecting Iraq's internet economy to reach USD $8.4 billion by 2028 — meaning the volume of Baghdadi and Basrawi commercial text requiring sentiment analysis will increase significantly over the coming years.

Case Study: Baghdad E-commerce Platform — Complaint Detection From 46% to 87% Recall

A Baghdad-based online retail platform — operating across Iraq with a large Basrawi and Baghdadi user base — needed a customer review sentiment system to process 38,000 monthly Arabic reviews and product ratings. Their existing model used an AraBERT fine-tuned on a pan-Arabic sentiment dataset that the vendor described as covering “all Arabic dialects” — but which, when inspected, contained under 2% Iraqi-origin text.

Before: The model achieved 57.4% overall sentiment accuracy on Iraqi dialect review text. Negative sentiment recall — the critical metric for product quality monitoring and supplier management — stood at 46.1%. Return-worthy product complaints were classified as neutral or positive at a rate of 31.7%, resulting in a 22-day average time between a complaint cluster forming and the operations team being alerted. Supplier retention penalties under the platform's SLA were triggered three times in six months due to missed complaint signals.

The annotation project delivered 24,000 labelled examples across positive, negative, neutral, and escalation-flag classes, with a sub-dialect breakdown of 14,000 Baghdadi, 7,000 Basrawi, and 3,000 Mosuli examples. Annotation was conducted by a team of twelve native Iraqi annotators — seven Baghdadi, three Basrawi, two Mosuli — with dialect-specific idiom tables covering Turkic loanword sentiment carriers, tribal complaint register, and Kurdish code-switching markers for the Mosuli pool. Final IAA kappa was 0.84 on primary sentiment classes and 0.77 on escalation flags.

After fine-tuning on the Iraqi-annotated dataset: Overall sentiment accuracy improved from 57.4% to 88.9%. Negative sentiment recall improved from 46.1% to 87.3%. Return-worthy complaint false negative rate fell from 31.7% to 6.2%. Supplier quality flagging response time dropped from 22 days to 3.1 days, eliminating the SLA penalty triggers. The platform identified four underperforming suppliers in the first two months post-deployment using signals the previous model had consistently missed.

The annotation project cost AUD $44,000 for annotation, QA, and delivery. The platform reported AUD $1.2M in supplier penalty avoidance and customer retention value in the 12 months following deployment.

Annotation Protocol Requirements for Iraqi Sentiment Projects

Iraqi sentiment annotation requires a structured protocol that differs from both generic Arabic NLP annotation and Khaleeji or Gulf-specific annotation approaches. The key requirements are:

Sub-dialect routing by source region. Text identified as Baghdadi, Basrawi, or Mosuli by dialect classifier or by platform region metadata should be routed to matching sub-dialect annotator pools. Mixing sub-dialects within a single project pool produces systematic IAA degradation on the religious-idiom and sarcasm categories. Even for national Iraqi platforms where sub-dialect routing is impractical at scale, annotators should be briefed on the sub-dialect distribution of the source corpus and given sub-dialect-specific idiom tables.

Turkic and Persian loanword sentiment tables. Annotation guidelines must include an Iraqi-specific vocabulary table covering the 20–30 most common Turkic and Persian loanwords used as sentiment carriers in Iraqi commercial text. These terms are absent from standard Arabic NLP guidelines and from the general Gulf Arabic idiom tables that some vendors repurpose for Iraqi projects. Without explicit coverage, Turkic-origin intensifiers and quality-judgement terms are routinely ignored or mis-scored.

Kurdish code-switching coverage for northern Iraqi text. Projects involving Mosuli or Kirkuki text — or national platforms where northern Iraqi user-generated content is a significant share — require annotators with Arabic-Kurdish bilingual competence, or a two-stage process where Kurdish-segment detection precedes sentiment labelling. Attempting to route Kurdish-embedded text exclusively to Arabic-monolingual annotators produces systematic coverage gaps on a category where the sentiment-carrying element is often the Kurdish segment.

Indirect complaint and tribal-register sarcasm flags. Iraqi annotation schema should include an explicit indirect-complaint flag (for tribal-register understatement) and a register-incongruity sarcasm flag (for Baghdadi formal-register sarcasm). These are annotation decisions that cannot be resolved by label distribution reweighting after the fact — they must be built into the schema from the start, with annotator training on the specific Iraqi constructions that trigger each flag.

Our Arabic NLP annotation service operates with dedicated Iraqi Mesopotamian annotator teams, including sub-dialect routing, Turkic vocabulary idiom tables, and Kurdish code-switching coverage for northern Iraqi text. Every project includes calibration pilots, two-stage QA, and IAA reporting at the sub-dialect level.

Related Reading

• Arabic Sentiment Analysis: The Complete Guide for MENA AI Teams
• Gulf (Khaleeji) Arabic Sentiment Analysis: What Models Get Wrong
• End-to-End Arabic Data Labeling: Project Case Study
• Arabic Data Labeling Service

Frequently Asked Questions