Iraqi Arabic Named Entity Recognition: What Models Get Wrong Without Native Annotators

Why Iraqi Arabic Is Uniquely Hard for Named Entity Recognition

Named entity recognition for Iraqi Mesopotamian Arabic sits at the intersection of three NLP challenges that no standard Arabic NER system was designed to handle simultaneously: the Turkic and Persian name layer from centuries of Ottoman administration and Persian cultural contact, the Kurdish language contact in northern Iraq that introduces Kurdish-origin entities into Arabic-language text, and the distinctive Iraqi tribal naming system that produces name spans structurally different from the Egyptian and Levantine naming patterns that anchor Arabic NER training data.

The root cause is dataset origin. Analysis of major Arabic NER corpora — ANERcorp, WikiFAIQ, the CAMeL NER dataset, and the SemEval 2022 Arabic NER track data — finds fewer than 4% Iraqi-origin entity instances (Pasha et al., 2014; Jarrar et al., 2023). Iraqi entity types are not merely underrepresented; they include structural categories absent from Egyptian, Levantine, and Gulf NER annotations — particularly tribal affiliations as name components and Kurdish-origin organisations as ORG entities. Research from the SemEval Arabic NER shared task and WANLP NER tracks documents 26–41% F1 degradation on Iraqi entity extraction for models trained on standard Arabic NER resources (Abdul-Mageed et al., 2023; Bouamor et al., 2019).

The practical stakes are significant. Iraq's banking and fintech sector is expanding rapidly under the Central Bank of Iraq's financial inclusion mandate, with mobile payment platforms and digital banking services creating large volumes of customer data requiring entity extraction for KYC and AML compliance. The World Bank's 2025 Iraq Financial Sector Development Report projects Iraq's formal banking population to grow from 23% to 45% of adults by 2028 — meaning the volume of Iraqi dialect KYC documents requiring NER is set to approximately double within three years.

Five Iraqi NER Failure Modes That Break Standard Arabic Models

1. Turkic-origin personal names misclassified as common nouns

A significant share of Iraqi personal names trace to Turkic roots through the Ottoman administrative period. Names like Nawroz (meaning New Year in Kurdish-Turkic), Dilnoza (heart delight — Turkic-Persian compound), Qasim (a name shared across Arabic but with distinct Iraqi sub-dialect frequency), and Firnas (Arabic-Turkic blend) are high-frequency in Iraqi name registers. Standard Arabic NER models trained on Egyptian and Levantine data — where these names are rare or absent — fail to recognise them as PER entities.

The failure mode is not simply out-of-vocabulary; it is active misclassification. Nawroz, for example, may be tokenised as a compound of common Arabic morphemes and assigned a non-entity tag. Qasim in Iraqi context frequently appears without the definite article or case marker that would help an MSA-trained model identify it as a proper noun. Non-Iraqi annotators who encounter these names in annotation tasks sometimes classify them as ORG entities (because they resemble certain Gulf institutional name patterns) or leave them untagged. Only annotators with Iraqi-region naming familiarity correctly and consistently classify the full Turkic-origin name inventory.

2. Kurdish entities in northern Iraqi text

Northern Iraqi Arabic-language text — from Mosul, Kirkuk, Erbil, and the Nineveh plain — contains Kurdish personal names, Kurdish organisation names, Kurdish place names, and hybrid Arabic-Kurdish entity spans that no standard Arabic NER model has training coverage for. Kurdish names in Arabic-language context are doubly challenging: they are not present in Arabic NER training corpora, and they do not follow Arabic morphological patterns that would allow rule-based entity detection to compensate.

For national Iraqi platforms — e-commerce, banking, government services — with northern Iraqi user bases, the Kurdish entity gap is a material accuracy problem. An e-commerce platform that cannot correctly extract the name “Peshtiwan Karim” as a PER entity from an order record is producing KYC and fraud-detection gaps on a portion of its user base. A document processing system that cannot identify “Kirkuk Provincial Council” (Majlis Muhafazat Kirkuk — often appearing with Kurdish-language acronyms) as an ORG entity is failing on a common northern Iraqi institutional reference. Fixing this requires annotators with Arabic-Kurdish bilingual competence, not simply Arabic-native annotators with Kurdish name lists.

3. Tribal naming conventions and multi-token personal name spans

Iraqi personal naming follows a patrilineal multi-generational structure combined with tribal affiliation markers that produces name spans structurally different from the patterns that anchor Arabic NER training data. A complete Iraqi formal name may include: a personal name, a father's given name, a grandfather's given name, and a tribal affiliation — for example “محمد حسن علي العبيدي” where العبيدي identifies tribal membership in the Ubaidi confederation. Standard NER models trained on Egyptian data — where family names in the European sense are common — consistently under-segment this span, treating العبيدي as a separate entity or as a non-entity adjective.

The span-boundary problem is not limited to four-token names. In formal Iraqi legal and administrative documents, full name renderings may run to five or six tokens. For KYC and AML applications, incorrect span boundaries produce entity mismatches against external name registries, triggering false positive alerts and manual review queues. Research on Arabic NER span-boundary errors finds that Iraqi name span detection has error rates 2.3–3.1 times higher than Egyptian name span detection when using the same Arabic NER model (Jarrar et al., 2023). The fix is annotator training on Iraqi tribal-name taxonomy and explicit span-boundary decision rules in annotation guidelines.

4. Location entity confusion in contested and renamed places

Northern Iraqi location entities present a specific NER challenge: many places have parallel Arabic and Kurdish names, some have been renamed since 2003, and some are contested between Iraqi government and Kurdistan Regional Government jurisdiction. A location like Kirkuk/كركوك may appear in Iraqi Arabic text by its Arabic name, its Kurdish name (Kerkûk), a historical Ottoman name, or an administrative designation that postdates the 2003 transition. Standard Arabic NER models trained on pre-2003 or non-Iraqi-sourced data have outdated or missing LOC entity coverage for the post-transition Iraqi location inventory.

For geospatial applications, logistics platforms, and government document processing systems operating across Iraq, the location entity coverage gap is directly operational. Routing and address verification systems that cannot normalise parallel Iraqi-Kurdish location names produce delivery failures. Compliance document extraction systems that cannot match post-2003 location designations against pre-2003 registry names produce KYC mismatches. Iraqi NER annotation guidelines must include an updated location taxonomy covering contested and renamed places, maintained by annotators who are familiar with post-2003 Iraqi administrative geography.

5. Post-2003 organisation entities and abbreviation patterns

The 2003 transition restructured the entire Iraqi institutional landscape, creating a new inventory of government ministries, political parties, armed forces components, and civil society organisations that do not appear in any pre-2003 Arabic NER training corpus and are underrepresented in post-2003 resources built primarily from non-Iraqi Arabic sources. Iraqi Arabic text frequently abbreviates these organisations using initialisms that do not follow Gulf or Egyptian abbreviation conventions.

The abbreviation challenge compounds with Iraqi Arabic orthographic variation: the same organisation may appear as a full name, an Arabic acronym, an English acronym (many post-2003 Iraqi organisations adopted English-language acronyms), or a colloquial Baghdadi nickname. Standard NER models trained on pre-2003 or non-Iraqi sources have no coverage for this entity inventory. Native Iraqi annotators recognise post-2003 Iraqi organisations immediately; non-Iraqi annotators — including Gulf Arabic annotators with no exposure to the post-2003 institutional landscape — require extensive reference documentation to annotate them correctly.

Iraqi NER vs Gulf and Egyptian NER: Why You Cannot Repurpose Gulf Annotators

A common vendor shortcut is to staff Iraqi Arabic NER projects with Gulf or Egyptian annotators — on the reasoning that Arabic is Arabic and entity types are entity types. The accuracy data does not support this. Comparative annotation studies using Gulf and Iraqi annotators on the same Iraqi-dialect NER corpus show IAA kappa degradation of 0.11–0.18 on personal name categories and 0.14–0.22 on location categories, driven primarily by the tribal naming span-boundary and Turkic-name classification failures documented above (Althobaiti & Albogami, 2022; Jarrar et al., 2023).

The entity category failure modes are distinct for different non-native annotator pools. Gulf Arabic annotators are familiar with Khaleeji personal name patterns but miss Turkic-origin Iraqi names and fail on tribal multi-token span boundaries. Egyptian annotators are better calibrated on personal name entity spans (Egyptian naming follows a similar multi-generational structure) but have no coverage for Kurdish entities or post-2003 Iraqi institutional names. Levantine annotators bring stronger familiarity with post-conflict institutional name changes but miss the Turkic name inventory and have low IAA on tribal affiliation spans.

None of these non-native pools handles the Kurdish entity problem at all. For Iraqi NER projects where northern Iraqi text is a material share of the corpus, Arabic-Kurdish bilingual annotators are a hard requirement — not an optional quality enhancement.

Case Study: Baghdad Fintech KYC Platform — Entity F1 From 51% to 88%

A Baghdad-based digital payments platform — processing KYC documents and transaction records for 180,000 Iraqi customers — needed named entity extraction to automate identity verification, AML screening, and customer record matching. Their existing system used multilingual BERT fine-tuned on a commercially sourced Arabic NER dataset described as covering “all major Arabic dialects” — but containing under 3% Iraqi-origin entity instances and no Kurdish entity coverage.

Before: The model achieved 51.3% micro F1 on Iraqi entity extraction across PER, ORG, LOC, and DATE categories on a held-out Iraqi dialect test set. PER entity F1 specifically stood at 44.7%, driven primarily by Turkic-origin name misclassification and tribal name span under-segmentation. KYC straight-through processing (where the NER output matched against external registry without manual review) was 18.4%. The remaining 81.6% of KYC documents required manual review, at a cost of AUD $3.20 per record. The Central Bank of Iraq's Compliance Division noted the platform in a supervisory visit for its low automation rate relative to its peer group.

The annotation project delivered 18,000 annotated sentences covering the full Iraqi entity taxonomy: PER (including Turkic-origin names and full tribal name spans), ORG (including post-2003 Iraqi institutional names and their abbreviation variants), LOC (including parallel Arabic-Kurdish location names and post-2003 administrative geography), and DATE (including Iraqi calendar systems used in legal documents). Annotation was conducted by a team of ten native Iraqi annotators — six Baghdadi, three Basrawi, one Mosuli Arabic-Kurdish bilingual — with a tribal-name taxonomy reference, a Turkic-name morphology table, and a post-2003 Iraqi institutional name directory built into the annotation guidelines. Final IAA kappa across entity categories was 0.87 on PER, 0.82 on ORG, 0.84 on LOC, and 0.91 on DATE.

After fine-tuning on the Iraqi-annotated dataset: Overall entity micro F1 improved from 51.3% to 88.1%. PER entity F1 improved from 44.7% to 85.3%. KYC straight-through processing increased from 18.4% to 71.2%. Manual review volume fell from 81.6% to 28.8% of KYC documents, reducing per-record manual review cost from AUD $3.20 to AUD $0.92. The Central Bank supervisory visit in the following quarter noted the platform's automation rate as now within the top quartile of licensed digital payment providers.

The annotation project cost AUD $39,000 for annotation, guideline development, QA, and delivery. The platform calculated AUD $2.4M in annualised manual review cost reduction from the improved KYC automation rate.

Annotation Protocol Requirements for Iraqi Arabic NER Projects

Iraqi Arabic NER annotation requires protocol elements that differ from generic Arabic NER and from Gulf or Egyptian dialect NER approaches. The key requirements are:

Tribal-name taxonomy in annotation guidelines. Guidelines must document the major Iraqi tribal confederations and their naming affiliation patterns — the Ubaidi, Shammar, Juburi, Dulaimi, and other major tribal groupings — so annotators can correctly identify tribal affiliation tokens as components of multi-token PER spans rather than as separate LOC or ORG entities. Without this documentation, tribal-affiliation token classification varies between annotators in ways that systematic calibration alone cannot resolve.

Turkic-name morphology table. A reference table of high-frequency Turkic-origin Iraqi personal names — with their morphological roots and frequency context — reduces misclassification of Turkic names as common nouns. The table does not need to be exhaustive; the 50–80 highest-frequency Turkic-origin names in Iraqi commercial and administrative text cover the majority of the classification problem.

Arabic-Kurdish bilingual coverage for northern Iraqi text. Projects with Mosuli, Kirkuki, or northern Iraqi text require at least one Arabic-Kurdish bilingual annotator or a two-stage protocol where Kurdish-origin entity spans are pre-identified before Arabic-native annotators classify entity type. Attempting to annotate Kurdish entities with Arabic-monolingual annotators produces systematic under-coverage even when Kurdish name lists are provided — because Kurdish code-switched spans require reading comprehension in Kurdish to correctly identify span boundaries.

Post-2003 institutional name directory. An updated Iraqi institutional name directory — covering ministries, security agencies, political parties, and civil society organisations post-2003 — enables consistent ORG entity classification for the current Iraqi institutional landscape. This directory should be maintained and updated as Iraqi institutional names continue to evolve; the 2021 Iraqi elections produced a new set of political coalition names that require updates to any pre-2021 Iraqi ORG entity reference.

Our Arabic NLP annotation service provides dedicated Iraqi Mesopotamian NER annotation with tribal-name taxonomy, Turkic-name morphology tables, Arabic-Kurdish bilingual annotator coverage, and post-2003 Iraqi institutional name directories. Every project includes calibration pilots, two-stage QA with entity-category IAA reporting, and delivery in CoNLL or JSON-Lines format.

Related Reading

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