Sudanese Arabic Dialect Identification: What Models Get Wrong Without Native Annotators

Why Pan-Arabic DID Models Cannot Classify Sudanese Arabic Varieties

Arabic dialect identification (DID) research has produced robust systems for the major Arabic dialect groups — Gulf, Egyptian, Levantine, Maghrebi, Iraqi — that are commercially relevant and well-represented in available labelled corpora. Sudanese Arabic, when it appears in DID training data at all, is typically assigned to a residual ‘other’ category or incorrectly mapped to Egyptian Arabic on the basis of geographic proximity and surface phonological similarity.

This categorisation error has consequences. Sudanese Arabic is not a phonologically simplified variety of Egyptian Arabic. It is a Nile Valley dialect with a distinct substrate (Nile Nubian languages), a distinct prestige layer (Classical Arabic phoneme retention patterns absent from Egyptian), and a distinct contact-language history (Nuba Mountains, Saharan, Chadic, and South Sudanese languages in its western and southern varieties). These features place Sudanese Arabic outside the Egyptian Arabic cluster in acoustic and lexical feature space — a misclassification that propagates downstream whenever DID is used to route Sudanese Arabic content.

Research on Arabic dialect variety differentiation using automated methods (Zaidan & Callison-Burch, 2014; Abdul-Mageed et al., 2020 PADIC extension) consistently documents performance degradation of 28–41% when pan-Arabic DID systems are evaluated on Nile Valley dialect test sets compared to their performance on the dialect groups they were trained to classify. For within-Sudan variety differentiation — distinguishing Khartoum from Northern Nile Valley from Kordofan from Darfur — automated systems achieve 41–56% F1, while native Sudanese listener judgements achieve 71–84% on the same ambiguous samples.

Our Arabic NLP annotation service provides Sudanese Arabic DID annotation across all five major regional varieties — Khartoum urban, Northern Nile Valley, Kordofan, Darfur, and Juba Arabic — with native annotators from each speech community and adjudicated review for convergence-zone ambiguous samples.

The Five Sudanese Arabic Varieties and Their Diagnostic Features

Khartoum urban Arabic

Khartoum urban Arabic is the prestige variety and the dominant digital register. It is characterised by relatively low Nubian lexical density compared to Northern varieties (most Nubian-origin terms have been replaced by MSA or Egyptian borrowings in urban speech), high rates of English sentence-level code-switching, a tendency toward glottal stop realisations for /q/ in casual speech, and significant loanword integration from international media sources. Khartoum urban Arabic is the most comprehensible variety to other Arabic speakers and the most likely to be underestimated in DID tasks — its MSA-adjacent surface features cause automated systems to classify it as Egyptian or MSA rather than Sudanese.

The diagnostic features for Khartoum urban Arabic in DID annotation are: specific Sudanese Arabic function word inventory (distinct from Egyptian and Gulf equivalents), characteristic Khartoum-specific discourse particles, English code-switching rate and pattern (sentence-level switches more frequent than word-level insertions), and absence of Northern Nile Valley Nubian lexical density.

Northern Nile Valley Arabic

Northern Nile Valley Arabic — spoken from the Egyptian border through Dongola, Karima, and Shendi to the Khartoum periphery — is the variety with the highest Nubian substrate lexical density. Nobiin and Dongolawi (Kenzi-Dongola) loanwords for agricultural vocabulary, body parts, kinship terms, and environmental features are abundant in this variety and substantially absent from Khartoum urban speech. Northern varieties also show the most consistent Classical Arabic /q/ retention as a uvular stop, the most pronounced Nubian phonological substrate features in consonant inventory and prosodic structure, and lower English code-switching rates than Khartoum urban.

For DID annotation, Northern Nile Valley Arabic is the most distinctive Sudanese variety in lexical feature space — the Nubian-origin vocabulary provides clear discriminating signals. However, the convergence zone between Northern Nile Valley and Khartoum urban (the Shendi–Khartoum North corridor) produces speakers with mixed feature profiles that require native annotators from both varieties to classify reliably.

Kordofan Arabic

Kordofan Arabic — spoken across the Nuba Mountains region, El Obeid, and the western Sudan interior — carries contact vocabulary from Nuba Mountains languages (a genetically diverse family including Niger-Congo and Nilo-Saharan branches). This contact layer is distinct from the Nile Nubian substrate of Northern varieties and from the Saharan/Chadic contact features of Darfur. Kordofan Arabic is the variety most likely to be misclassified by automated systems — its lexical features overlap partially with Khartoum urban (which has absorbed some Nuba Mountains vocabulary through urbanisation) and partially with Darfur (which also shows non-Arabic substrate contact vocabulary).

Annotation of Kordofan Arabic for DID requires annotators specifically from the Kordofan region or with extended linguistic exposure to the variety. Khartoum urban-native annotators cannot reliably distinguish Kordofan contact vocabulary from Darfur contact vocabulary, and Northern Nile Valley annotators cannot reliably distinguish Kordofan features from the mixed urban periphery varieties.

Darfur Arabic

Darfur Arabic shows the influence of Saharan and Chadic language contact — from the Zaghawa, Fur, Masalit, and other Darfurian communities. This produces diagnostic vocabulary from Nilo-Saharan and Afroasiatic (Chadic) contact that is absent from other Sudanese Arabic varieties. Darfur Arabic also shows distinctive prosodic features attributable to the tonal languages in contact with Arabic in western Sudan. In digital contexts, Darfur Arabic is underrepresented relative to its speaker population — conflict-related displacement and limited digital infrastructure in western Sudan mean that Darfur Arabic digital text is sparse.

DID annotation for Darfur Arabic requires annotators with Darfur-region background. The conflict history — including post-2003 and post-2023 displacement — means that many Darfur-variety native speakers are now located in IDP camps, refugee communities in Chad, or diaspora communities in Europe and North America. Diaspora Darfur Arabic annotators are a viable source if their variety exposure pre-dates heavy dialect levelling.

Juba Arabic

Juba Arabic is the most linguistically distinctive of the Sudanese Arabic varieties — it is a contact language (creole or advanced pidgin, depending on the linguistic analysis) that developed as a lingua franca in the Equatoria region of what is now South Sudan, with heavy substrate influence from Nilotic and Central Sudanic languages. It uses Arabic-derived vocabulary restructured by a non-Arabic grammatical system, with drastically simplified nominal and verbal morphology compared to any other Arabic variety.

For DID purposes, Juba Arabic is the easiest Sudanese variety to classify by automated means — its simplified morphology and distinctive Nilotic-substrate vocabulary produce feature profiles far from any other Arabic variety. However, for downstream annotation tasks, Juba Arabic requires dedicated annotators who are native or highly proficient in the variety — an MSA-proficient Arabic annotator cannot reliably annotate Juba Arabic NER, sentiment, or intent data, even though the vocabulary is predominantly Arabic-derived.

Case Study: Khartoum Media Platform — Dialect Routing Accuracy 48% to 81%

A Khartoum-based digital media platform serving national Sudanese audiences needed dialect routing to direct user-generated content to correctly-matched downstream annotation pipelines — sending Khartoum urban text to the urban annotator pool, Northern Nile Valley content to Northern annotators, and so on. The platform was producing labelled datasets for a Sudanese Arabic language model pre-training project. Incorrect dialect routing meant that Northern Nile Valley content was being annotated by Khartoum-background annotators who did not recognise Nubian-substrate vocabulary, producing NER and sentiment labels with systematic errors on regionally-specific vocabulary.

Before: The platform's initial DID system used a pan-Arabic classifier fine-tuned on a small Egyptian and Gulf Arabic dataset with a Sudanese Arabic category populated from social media posts labelled by Khartoum-only annotators. This system achieved 48.3% overall accuracy across the five Sudanese regional varieties. For Northern Nile Valley content specifically, accuracy was 31.6% — more than two-thirds of Northern content was being routed to the wrong annotator pool. Kordofan accuracy was 38.2%, Darfur 41.7%. Juba Arabic was the exception at 71.4%, due to its distinctive grammatical features.

The annotation project labelled 6,400 Sudanese Arabic text samples across the five varieties — 1,800 Khartoum urban, 1,400 Northern Nile Valley, 1,100 Kordofan, 900 Darfur, 1,200 Juba Arabic. Each sample was labelled by two independent native annotators from the relevant variety, with adjudication by a third annotator for the 14.3% of items where the two first-pass annotators disagreed. The annotation team comprised eleven native Sudanese annotators covering all five regional varieties, with a senior Sudanese linguist serving as adjudicator and taxonomy custodian. Inter-annotator kappa on the five-class task was 0.76 overall — 0.81 for Khartoum and Juba (most distinctive), 0.68 for the Northern/Kordofan convergence zone (most ambiguous).

After fine-tuning the DID classifier on the labelled dataset: Overall five-class accuracy improved from 48.3% to 81.2%. Northern Nile Valley accuracy improved from 31.6% to 74.8% — still the lowest-performing class due to convergence-zone ambiguity, but now above the routing-reliability threshold. Kordofan improved from 38.2% to 77.3%, Darfur from 41.7% to 79.6%. Juba Arabic improved from 71.4% to 91.8%. Downstream annotation error rates on regionally-specific vocabulary fell by 61% across all five annotation pipelines within 60 days of routing system deployment, reducing the relabelling workload by an estimated 1,100 items per month. The annotation project total cost was AUD $29,700; the platform attributed AUD $18,400 per month in saved relabelling and annotation rework costs.

Convergence Zones and Ambiguous DID Samples

The most challenging DID annotation in Sudanese Arabic occurs at convergence zones — geographic and sociolinguistic transition areas where speaker feature profiles blend two regional varieties. The Khartoum–Northern Nile Valley convergence zone (Shendi, Khartoum North, Omdurman peri-urban areas) produces speakers who retain Northern Nubian vocabulary in specific semantic domains (agricultural and kinship vocabulary) while using Khartoum urban features in technology and commercial domains.

These convergence-zone speakers produce content that native annotators from either variety can identify as ‘somewhere between Khartoum and Northern’ but cannot reliably assign to a single class. The appropriate annotation protocol for convergence-zone samples is a multi-label approach — assigning probability weights across the two most plausible classes rather than forcing a single label — or an explicit convergence-zone class that routes content to a mixed annotator pool.

A similar convergence zone exists between Kordofan and Darfur Arabic in the border areas of South Kordofan and Darfur, where Nuba Mountains and Saharan contact vocabulary co-occur in the same speaker repertoire. Automated DID systems perform particularly poorly here; native annotator performance is only marginally better, with kappa around 0.62–0.68 for the Kordofan/Darfur boundary task.

For a broader treatment of annotation quality metrics and IAA interpretation, see our Cohen's kappa annotation quality guide, which covers when lower IAA reflects genuine category ambiguity rather than annotator error.

Practical Applications for Sudanese Arabic DID

Annotator routing for downstream tasks. The primary commercial application of Sudanese Arabic DID is routing text or audio samples to the correctly-matched annotator pool for NER, sentiment, intent, or moderation tasks. Routing Northern Nile Valley content to Khartoum urban annotators produces systematic errors on Nubian-substrate vocabulary; routing Kordofan content to Northern annotators produces errors on Nuba Mountains contact vocabulary. Accurate DID-based routing improves downstream annotation quality by 25–40% on regionally-specific vocabulary categories, based on the case study results described above.

ASR acoustic model selection. Sudanese Arabic ASR model performance varies significantly across regional varieties. A Khartoum urban-adapted acoustic model achieves lower WER on Northern Nile Valley speech than an unadapted MSA model, but a Northern-adapted model achieves 20–30% lower WER than the Khartoum-adapted model on Northern Nile Valley test sets. DID-based ASR routing — classifying incoming Sudanese Arabic audio by regional variety before selecting the acoustic model — provides measurable WER reductions over single-model deployment. See our post on Sudanese Arabic speech transcription annotation for the full ASR pipeline context.

Regional personalisation. Digital platforms serving national Sudanese audiences can use DID to personalise content recommendations, advertising, and interface language to regional variety preferences. Sudanese users show measurable engagement differences across regional variety content — Khartoum urban content performs less well with Northern Nile Valley-variety users when it contains urban vocabulary absent from their active lexicon. DID-based regional personalisation requires a DID classifier validated on the full variety range, not just Khartoum urban vs. other.

Sociolinguistic research and language documentation. The Nile Valley Arabic varieties — particularly Northern Nile Valley Arabic with its Nubian substrate — are at risk of vocabulary erosion through urbanisation and Khartoum urban influence. Large-scale DID annotation corpora provide documentary resources for language preservation research, in addition to their commercial applications.

For guidance on our full Arabic annotation pipeline, see our Arabic NLP annotation service and our post on building custom Arabic NLP datasets.

Related Reading

• Sudanese Arabic Content Moderation: What Models Get Wrong Without Native Annotators
• Sudanese Arabic Named Entity Recognition: What Models Get Wrong Without Native Annotators
• Gulf (Khaleeji) Arabic Dialect Identification: What Models Get Wrong Without Native Annotators
• Arabic NLP Annotation Service

Frequently Asked Questions