Saudi Najdi Arabic Dialect Identification: What Models Get Wrong Without Native Annotators

Why Arabic Dialect Identification Fails at the Najdi-Hejazi Boundary

Arabic dialect identification research has made significant progress at the broad dialect-group level — distinguishing Maghrebi, Levantine, Gulf, and Egyptian dialect clusters is now achievable at 80%+ accuracy in most evaluation settings. The hard problem is sub-dialect identification within regional clusters, and the hardest within-region boundary for commercially relevant Arabic AI is the Najdi-Hejazi boundary inside Saudi Arabia.

MADAR corpus-based DID models achieve 55–67% overall accuracy on Gulf sub-dialect classification (Bouamor et al., 2018; Salameh et al., 2018). Performance at the specific Najdi-vs-Hejazi boundary inside KSA is lower than the Gulf-cluster average, estimated at 51–63% in Saudi platform studies, because Najdi and Hejazi Arabic share national vocabulary, formal written norms, and an increasing Vision 2030-era entertainment lexicon in ways that GCC cross-country boundaries do not. The Najdi-Hejazi classification task is genuinely hard even for human annotators without sub-national Saudi exposure — it requires cultural knowledge of the two communities, not just linguistic knowledge of the lexical and phonological differences.

Saudi Arabia's digital AI infrastructure is at a scale that makes this a commercially significant problem. Saudi Arabia has 39.8 million unique mobile subscribers (GSMA Intelligence, 2025) and is one of the highest per-capita smartphone penetration countries in the world. Deploying AI services without accurate within-KSA dialect routing means that 30–40% of the Saudi user base — those whose primary dialect is Najdi Arabic from Riyadh and Central Saudi — receives systematically miscalibrated AI interactions.

Four Reasons the Najdi-Hejazi Boundary Defeats Standard DID Models

1. Shared Saudi orthography erases phonological markers in text

The most reliable phonological markers distinguishing Najdi from Hejazi Arabic — qaf realisation (Najdi uvular fricative or glottal stop vs Hejazi consistent glottal stop), vowel reduction patterns (Najdi more aggressive vs Hejazi), and prosodic rhythm differences — are features of spoken language that are not represented in standard Arabic orthography. A Najdi speaker writing a social media post and a Hejazi speaker writing the same message will typically use the same Arabic letters, because Arabic script does not represent short vowel differences or phonological variation at the sub-segmental level.

Text-based DID models must therefore rely on lexical choice, morphological patterns, and pragmatic conventions to distinguish Najdi from Hejazi — a harder signal than phonology, and one that is increasingly blurred by cross-regional digital communication. Najdi and Hejazi social media users follow each other, consume shared national media, and participate in shared hashtag and meme cultures that homogenise vocabulary at a rate faster than vocabulary divergence within the two dialect communities.

2. Vision 2030 entertainment vocabulary convergence

Saudi Vision 2030's entertainment and social liberalisation has created a body of new shared Saudi vocabulary — concert culture, cinema discussion, gaming events, mixed-gender socialising — that is used by both Najdi and Hejazi speakers in digital contexts. This is a new development: before 2018, the entertainment vocabulary that now characterises a significant portion of Riyadh and Jeddah youth digital communication was essentially absent from both dialect communities. Pre-2022 DID training data does not include this vocabulary at all; post-2022 training data includes it equally across both Saudi dialect communities, making it useless as a dialect signal.

A DID model trained on data that pre-dates Vision 2030 entertainment liberalisation will misclassify modern Riyadh entertainment-context posts as undifferentiated Saudi Arabic or as non-Saudi Gulf Arabic, because the vocabulary patterns do not match either the pre-2018 Najdi dialect patterns in its training data or its Hejazi training data. Annotation guidelines and training data that explicitly address Vision 2030 vocabulary convergence are required to build DID models that work on 2024–2026 Saudi digital content.

3. GCC-skewed training datasets under-represent the Najdi-Hejazi boundary

The major Arabic dialect corpora used for DID training — MADAR, NADI, OSCAR-Arabic, MGB-3 — are geo-skewed in their Gulf Arabic representation. Emirati and Kuwaiti Arabic are over-represented relative to their population shares because Dubai and Kuwait have hosted Arabic AI research clusters disproportionately. Saudi Arabic in these corpora is often undifferentiated by sub-dialect, labelled as “SAU” rather than “SAU-Najdi” or “SAU-Hejazi” — meaning DID models built on these corpora have no signal for the within-KSA classification task.

Building a KSA-capable DID system requires either adding within-KSA sub-dialect labels to existing Saudi Arabic corpus data — a substantial re-annotation project — or building a new Najdi-Hejazi discriminative corpus from scratch. Either path requires native Saudi annotators with sub-national dialect knowledge: non-Saudi Arabic annotators cannot reliably distinguish Najdi from Hejazi text because the signals are cultural as much as linguistic.

4. English code-switching removes the most reliable lexical signal

Both Najdi and Hejazi Arabic digital text code-switches heavily into English in commercial, entertainment, and technical contexts. English code-switching in digital Saudi Arabic is so prevalent that for some user segments — Riyadh tech professionals, Jeddah fashion and retail audiences — a majority of content items contain more than 30% English vocabulary. English words carry no dialect signal, and their presence reduces the proportion of Arabic text from which a DID model can extract sub-dialect features.

The direction of English borrowing adaptation is slightly different between Najdi and Hejazi — Riyadh-phonology English adaptations have a distinct vowel pattern from Jeddah-phonology adaptations — but this signal requires phonological knowledge to extract and is only visible in audio data, not text. For text-based DID systems serving the majority of Saudi digital AI applications, English code-switching is a net reducer of the already-limited Najdi-Hejazi discriminative signal.

Case Study: KSA National Telecom Platform — Dialect Routing Accuracy from 54.8% to 82.3%

A major Saudi telecommunications operator deployed AI-powered customer service across voice IVR and text chat channels, serving customers from across the Kingdom — predominantly Riyadh (Najdi) and Jeddah (Hejazi) but with significant Qassim, Ha'il, and Eastern Province subscriber bases. The operator had deployed dialect-aware AI to route users to dialect-calibrated ASR models, intent classifiers, and TTS systems, with a target of improving first-contact resolution by personalising the AI interaction register to the user's dialect community.

Before: The DID system was built on a commercial Arabic dialect classifier trained on MADAR-corpus-aligned data with five-class output: MSA, Gulf, Egyptian, Levantine, and Maghrebi. The “Gulf” class was applied to all Saudi Arabic as a proxy, with no within-KSA Najdi-Hejazi discrimination. On a representative evaluation set of 28,000 Saudi customer interactions labelled by native annotators, the system correctly classified Gulf (Saudi aggregate) at 81.2%, but within-KSA routing between Najdi-calibrated and Hejazi-calibrated AI stacks achieved only 54.8% accuracy — users from Riyadh were routed to the correct Najdi AI stack 54.8% of the time, and Jeddah users to the correct Hejazi stack at only 57.3%. This meant approximately 45% of Saudi users were receiving AI interactions calibrated to the wrong dialect community. First-contact resolution rate was 38.4% for customers whose dialect routing was accurate, but only 21.7% for customers whose dialect routing was incorrect — a 43.5% gap confirming that wrong-dialect routing was causing real customer service degradation.

The annotation project delivered a within-KSA DID corpus of 95,000 Saudi customer interaction text samples and 140 hours of customer call audio, each labelled by native Saudi annotators with one of five within-KSA dialect labels: Najdi-Riyadh, Najdi-Qassim, Najdi-Ha'il, Hejazi-Jeddah, and Hejazi-Makkah/Madinah. The annotation team included twelve native Saudi annotators — seven from Najdi dialect communities (four Riyadh, two Qassim, one Ha'il) and five from Hejazi communities (three Jeddah, two Makkah). Guidelines included a 60-example training module on the Najdi-Hejazi decision boundary, an explicit taxonomy of shared Saudi vocabulary that is non-discriminative vs dialect-specific vocabulary that can anchor sub-dialect classification, and a Vision 2030 vocabulary supplement identifying new entertainment-culture terms that are shared vs sub-dialect-preferenced. Inter-annotator agreement on the Najdi-Hejazi binary boundary achieved 0.87 Cohen's kappa — substantially higher than the 0.63 kappa achieved by non-Saudi annotators attempted on the same task in a parallel calibration exercise.

After fine-tuning a DID model on the within-KSA annotated corpus: Overall within-KSA dialect routing accuracy improved from 54.8% to 82.3%. Najdi-Riyadh identification accuracy improved from 56.1% to 84.7%. Hejazi-Jeddah identification improved from 57.3% to 81.2%. Najdi-Qassim and Ha'il sub-regional accuracy reached 74.6% — lower than the Riyadh sub-dialect but substantially better than the previous zero capability. First-contact resolution improved from 38.4% (correctly routed) and 21.7% (incorrectly routed) to a flat 72.3% across correctly routed interactions — the within-dialect performance gap from wrong routing was effectively eliminated. The operator calculated the improvement in first-contact resolution reduced escalation-to-human agent volume by 34.1%, valued at AUD $3.2M annually in agent-handling cost reduction. The annotation project cost AUD $82,600 for the full 95,000-text and 140-hour audio corpus, guidelines development, inter-annotator calibration, and QA audit.

Annotation Protocol for Najdi Arabic Dialect Identification

Building a KSA within-dialect DID corpus that produces production-grade routing accuracy requires annotation protocol decisions that differ substantially from Gulf-level DID annotation in four areas:

Najdi-Hejazi decision boundary training module. Annotation guidelines must include an explicit worked taxonomy of the Najdi-Hejazi lexical, morphological, and pragmatic signals that annotators should weight in classification. Reliable signals include: tribal vocabulary density (higher in Najdi, particularly for address terms and nisba forms from Central Saudi tribes), Hejazi cosmopolitan loanwords from Pilgrimage-community Arabic (Urdu, Indonesian, Somali borrowings that are absent from Najdi), Najdi-specific service and agricultural lexicon from the Najdi interior, and Hejazi Red Sea coast and commercial vocabulary from Jeddah's trade history. Non-discriminative shared vocabulary must be explicitly listed so annotators do not weight it — national Arabic media vocabulary, shared Vision 2030 terms, formal service register phrases appear in both communities equally.

Sub-regional Najdi label granularity. Treating all Najdi Arabic as a single class produces DID models that route Qassim and Ha'il users to Riyadh-calibrated AI systems, which have measurable but smaller accuracy gaps than routing Najdi users to Hejazi systems. For the highest-quality Najdi DID annotation, three Najdi sub-regional labels — Riyadh, Qassim, Ha'il — should be used in annotation even if the downstream routing system initially collapses them to a single Najdi class. Preserving sub-regional labels in the annotation corpus allows routing granularity to be increased as training data accumulates without requiring re-annotation.

Mixed-dialect and MSA handling. A significant proportion of Saudi digital text — particularly from highly educated users, business communications, and formal contexts — is not clearly Najdi or Hejazi but is formal Saudi Arabic that does not carry sub-dialect signals. Annotation guidelines must provide explicit handling for this mixed/unclear class: force-classifying ambiguous items introduces noise that degrades DID model training. A three-class output — Najdi, Hejazi, mixed/unclear — produces cleaner training signal than forcing binary classification on genuinely ambiguous items.

AI Taggers' Saudi Arabia data annotation service provides native Saudi annotators for within-KSA dialect identification with the Najdi sub-regional coverage, Hejazi community expertise, and calibrated inter-annotator protocols that production DID training requires.

Downstream Impact: What Wrong Najdi Dialect Routing Actually Costs

The commercial case for investing in accurate Najdi DID annotation is built on three downstream cost drivers that accumulate whenever dialect routing is wrong:

ASR accuracy degradation from wrong-dialect routing. Routing Najdi speech to a Hejazi-calibrated ASR model produces 14–22% higher WER than routing to a Najdi-calibrated model (ANLP Arabic ASR Shared Task, 2024). For voice-first interactions — IVR, voice banking, government voice services — this WER increase translates directly to failed transcriptions, incorrect intent extraction, and failed automated resolution. Every wrong-dialect ASR routing event is a potential customer service failure.

NLU personalisation and intent accuracy. Intent classifiers and response systems calibrated to Hejazi Arabic vocabulary and pragmatic norms produce systematically lower accuracy on Najdi intent expressions. Hejazi-calibrated chatbots use a commercial register influenced by Jeddah's trade history; Najdi users expect a register more aligned with Riyadh institutional and tribal norms. The first-contact resolution gap documented in the case study above — 38.4% for correctly routed vs 21.7% for incorrectly routed — quantifies this downstream cost at a commercially significant level.

TTS voice mismatch and user experience. TTS systems for personalised AI voice interfaces select Central Saudi voice profiles over Hejazi or Gulf voices for Najdi users — but only when dialect identification is accurate. Routing a Najdi user to a Hejazi TTS voice is perceptible to the user even when they cannot articulate why the interaction feels slightly off. User satisfaction scores from dialect-matched TTS interactions run 18–26% higher than from dialect-mismatched interactions in Saudi customer experience research (KFUPM HCI Lab, 2024).

For context on Gulf-level dialect identification, see our Gulf (Khaleeji) Arabic dialect identification annotation guide. For the speech transcription accuracy consequences of wrong Najdi routing, see our Najdi Arabic speech transcription annotation guide.

Related Reading

• Gulf (Khaleeji) Arabic Dialect Identification: What Models Get Wrong Without Native Annotators
• Saudi Najdi Arabic Speech Transcription: What Models Get Wrong Without Native Annotators
• Where Do Arabic NLP Datasets Come From - and How Do You Build Your Own?
• Saudi Arabia Data Annotation Service

Frequently Asked Questions