Saudi Najdi Arabic Sentiment Analysis: What Models Get Wrong Without Native Annotators

What Makes Najdi Arabic a Distinct Sentiment Problem

Najdi Arabic is the dialect spoken across Central Saudi Arabia — Riyadh, Qassim, Ha'il, and the surrounding Najd plateau. It is the dominant register in KSA government, enterprise, and major e-commerce platforms, which means any AI product deployed for the Saudi market at scale will encounter Najdi text. Yet Najdi is one of the least-represented Arabic dialects in academic NLP resources, and it is structurally distinct from the Egyptian and Levantine data that anchor most Arabic sentiment benchmarks.

The fundamental problem is dataset origin. The major Arabic sentiment resources — LABR (Large-scale Arabic Book Reviews), ASTD (Arabic Sentiment Twitter Dataset), SemEval Arabic subtasks — draw overwhelmingly from Egyptian, Levantine, and Gulf coastal sources. When these corpora do include Saudi text, it skews toward Hejazi Arabic (Jeddah, Makkah) rather than Najdi. The result is that Central Saudi dialect vocabulary, idiom, and pragmatic convention are essentially absent from the training data of standard Arabic sentiment models.

Research presented at WANLP (Workshop on Arabic Natural Language Processing) and ACL's Arabic NLP track documents the scale of the problem: Arabic sentiment models trained on MSA-dominant or non-Gulf data show 30–42% accuracy degradation on Central Saudi dialect test sets, with the worst performance on informal Najdi social media and review text (Al-Twairesh et al., 2021; Abdul-Mageed et al., 2023). Negative sentiment recall is the category that fails most severely, because Najdi complaint register relies on indirect construction rather than explicit negative vocabulary.

Five Najdi Sentiment Patterns That Break Arabic Models

1. Conservative complaint understatement

Najdi social register places a high cultural value on face-saving in public expression. Strong negative sentiment — particularly about products or services — is routinely expressed through understatement rather than explicit criticism. “ما عجبني كثير” (“it didn't please me much”) signals a serious complaint in Najdi context but reads as mild dissatisfaction to an MSA classifier. “يمكن يتحسن” (“maybe it will improve”) is often a polite way of saying “this is not good enough” — a construction that routes to neutral in MSA-trained models but to strongly negative in native Najdi annotation.

This pattern is more pronounced in Najdi than in other Gulf dialects. Emirati and Kuwaiti Arabic, which are more cosmopolitan in origin, have higher rates of direct negative expression. Najdi understatement is systematically missed by any sentiment model not explicitly trained on Central Saudi text, and it cannot be fixed by reweighting the label distribution — it requires annotators who recognise the indirection from within the cultural register.

2. Tribal and religious register as sentiment carrier

Najdi Arabic carries strong tribal social register — references to honour, generosity, and group belonging that encode sentiment in ways non-Najdi readers miss. A positive review that praises a business using tribal hospitality metaphors (“خدمة الكرماء” — “service of the generous”) signals deep positive sentiment that a non-tribal-context reader might treat as merely polite. Similarly, comparisons involving tribal dishonour (“تصرف ما يليق” — “behaviour that is not befitting”) carry strong negative weight in Najdi context but are often scored as neutral by standard models.

Religious idioms function as sentiment amplifiers in specific Najdi ways. “والله” (by God) used as a sentiment intensifier before a negative statement amplifies the complaint in Najdi social register. “يُسعدك” (“may you be made happy”) can be a sincere positive, a polite transition, or a veiled negative depending on what precedes and follows it — a distinction a native Najdi speaker makes immediately, and that no MSA model currently resolves correctly.

3. Najdi-specific vocabulary absent from MSA corpora

Najdi Arabic has distinct lexical items for quality judgements that do not appear in MSA training data. “مرة” (literally “a time/very”) functions as an intensifier — “مرة زين” means “very good”. “ما فيه” (“there is nothing to it”) signals negative quality judgement. “شكل” used informally means “it seems/looks like” as an epistemic hedge that modifies sentiment certainty. None of these usages reliably appear in Arabic sentiment resources built from news, Egyptian social media, or Levantine review data.

Contemporary Najdi urban Arabic — particularly Riyadh youth register — has also developed slang terms for quality evaluation that postdate existing NLP resources. “روعة” as a sentiment marker for “outstanding”, “مسوي” in commercial negative contexts, and code-switching with English terms (“الـ quality كانت تعبانة” — “the quality was poor”) create parsing challenges for any system not explicitly trained on contemporary Najdi text.

4. Sarcasm through exaggerated deference

Najdi sarcasm often takes the form of exaggerated deference rather than hyperbolic praise. A complaint framed as “بارك الله فيكم على الخدمة الرائعة” (“may God bless you for this wonderful service”) followed by a description of a failed experience is an established Najdi sarcastic construction. The religious blessing phrase triggers positive classification in MSA models; the native Najdi reader identifies the sarcasm from the structural mismatch between the deferential opening and the complaint that follows.

This construction is distinct from the Emirati exaggerated-praise sarcasm pattern documented in Khaleeji-level research. Najdi sarcasm through deference requires annotators who understand Central Saudi conversational norms, not simply Gulf Arabic pragmatics. The two sub-dialects handle sarcasm through different structural patterns, and confusing Khaleeji-generic annotation with Najdi-specific annotation produces lower IAA on precisely this category.

5. Code-switching patterns unique to Riyadh's commercial register

Riyadh's commercial and business text has specific Arabic-English code-switching patterns driven by Saudi Vision 2030's economic diversification. Reviews and social media posts about sectors transformed by Vision 2030 — entertainment, financial services, tourism, technology — show higher English code-switching rates than traditional sectors. The sentiment-carrying element can be in either language, and the entity being evaluated is often an English brand name embedded in an Arabic sentiment clause. Current Arabic sentiment models handle this switching inconsistently because their training data predates the post-2016 commercial expansion that drove the code-switching increase.

Najdi vs Hejazi: Why You Cannot Use One Saudi Annotator Pool for All KSA

Saudi Arabia contains two dominant Arabic dialect regions with meaningfully different sentiment expression patterns: Najdi (Central Saudi, Riyadh) and Hejazi (Western Saudi, Jeddah and Makkah). Many annotation vendors treat “Saudi Arabic” as a single category and staff projects with whoever is available — resulting in Hejazi annotators labelling Najdi text, or vice versa.

The differences matter for sentiment. Hejazi Arabic, shaped by Jeddah's port-city cosmopolitanism and higher exposure to Egyptian and Levantine Arabic through media and commerce, has more direct negative expression than Najdi. A Hejazi annotator will score some Najdi understatements as neutral when a Najdi-native annotator would score them as negative, because the Hejazi register is not trained in the same face-saving complaint indirection. The IAA degradation from cross-dialect annotation in Saudi Arabic sentiment projects has been measured at 0.08–0.14 kappa points on negative sentiment categories (Althobaiti & Albogami, 2022).

For brands operating primarily in Riyadh or targeting KSA government and enterprise markets — where Najdi is the dominant register — a Najdi-native annotator pool is not optional. It is a quality requirement. The Saudi AI market is projected to reach USD $9.7 billion by 2030 (IDC Saudi Arabia AI Market Report, 2025), with the vast majority of KSA enterprise AI deployments concentrated in Riyadh. Products that cannot accurately interpret Najdi sentiment are building on a broken foundation.

Case Study: Riyadh Fintech Platform — Najdi Complaint Detection From 44% to 89% Recall

A Riyadh-based digital banking platform needed a customer feedback sentiment system to process 55,000 monthly Arabic reviews, support tickets, and in-app survey responses. Their existing system used an AraBERT model fine-tuned on a commercially sourced Arabic sentiment dataset that its vendor described as “Saudi and Gulf Arabic” — but which, on inspection, contained primarily Hejazi and pan-Gulf coastal text.

Before: The model achieved 61.3% overall sentiment accuracy on Najdi dialect customer text. Negative sentiment recall — the most commercially important metric for a financial services complaints workflow — stood at 44.2%. Escalation-worthy complaints were classified as neutral at a rate of 28.4%, resulting in a 19-day average time-to-resolution for issues that should have been flagged within 24 hours. SAMA (Saudi Central Bank) conducted a service quality inquiry noting the platform's slow complaint response rate.

The annotation project delivered 28,000 labelled examples across positive, negative, neutral, and escalation-flag classes, sourced entirely from the platform's own Najdi user base. Annotation was conducted by a team of ten native Najdi annotators — eight from Riyadh, two from Qassim — with a two-stage QA protocol including expert adjudication on any item where the primary annotator flagged Najdi-specific idiom. Final IAA kappa across the full dataset was 0.86 on primary sentiment classes and 0.79 on the escalation flag class.

After fine-tuning on the Najdi-annotated dataset: Overall sentiment accuracy improved from 61.3% to 89.7%. Negative sentiment recall improved from 44.2% to 88.6%. The escalation-flag false negative rate fell from 28.4% to 5.1%. Average time-to-resolution for complaint tickets dropped from 19 days to 2.8 days as escalation routing became reliable. SAMA acknowledged the improvement in a subsequent service quality review.

The annotation project cost AUD $52,000 for annotation, QA, and delivery. The platform calculated AUD $1.8M in regulatory risk mitigation and customer retention value from the improved complaint handling in the 12 months following deployment.

Annotation Protocol Requirements for Najdi Sentiment Projects

Najdi sentiment annotation requires a structured protocol that differs from both generic Arabic NLP annotation and Khaleeji-level GCC annotation. The key requirements are:

Najdi-native annotator routing. Source text identified as Najdi — by dialect classifier or manual inspection — must go exclusively to annotators who are native Central Saudi speakers. Hejazi annotators should be separately routed to Hejazi-dialect text. Mixing annotator sub-dialects within a single Najdi sentiment project produces systematic IAA degradation on the complaint-understatement categories that matter most.

Explicit escalation and sarcasm flags. Standard three-class sentiment is insufficient for commercial Najdi text. Adding an escalation flag (for customer service workflows) and a sarcasm marker (for brand monitoring and social listening) reduces the mislabelling of Najdi indirect complaint and deference-based sarcasm. These are not post-processing additions — they must be built into the annotation schema from the start.

Religious idiom tables in annotation guidelines. Annotation guidelines must include a Central Saudi-specific idiom table — separate from a general Gulf Arabic idiom table — covering the 15–20 most common Najdi religious and tribal idioms used as sentiment carriers. Generic Khaleeji idiom guides underrepresent Najdi-specific usage and do not cover the tribal-register markers that are distinctive to Central Saudi sentiment expression.

Multi-annotator adjudication on indirect expressions. Items where the primary Najdi annotator flags sentiment ambiguity should go to a second Najdi-native annotator, not to a project supervisor from a different dialect. The disambiguation must happen within the Najdi register, not across it. Items that reach genuine two-annotator disagreement after Najdi adjudication can be escalated to expert review — but that should be a small fraction (typically 2–5%) of the full dataset.

AI Taggers' Saudi Arabia annotation service operates with dedicated Najdi-native annotator teams for Central Saudi projects, including the dialect-specific guideline development, calibration, and IAA reporting that KSA enterprise AI deployments require.

PDPL Compliance for Najdi Sentiment Annotation

Saudi PDPL (Personal Data Protection Law), administered by SDAIA, applies to Najdi sentiment annotation projects whenever source data contains personal data of Saudi individuals. Customer reviews, social media posts, and call-centre transcripts that include identifiable names, account references, or location-identifying phrases are in scope. The fact that data comes from a Saudi user base — as is almost invariably the case for Najdi-dialect annotation projects — does not change the PDPL applicability; it increases the need for PDPL-compliant annotation workflow design.

The practical compliance steps for Najdi sentiment annotation are: de-identify all source text before it leaves the production environment; remove names, phone numbers, and account identifiers that could link a text to an individual; establish data transfer agreements with annotation vendors specifying KSA data residency and access logging requirements; and maintain a data processing record that documents the basis for processing, the categories of data processed, and the annotation workflow controls in place.

For financial services platforms — such as the case study above — SAMA service quality obligations interact with PDPL, meaning that complaint-processing workflows must both protect personal data and respond to complaints within mandated timeframes. The annotation system underpinning the complaint classification must be built to both standards simultaneously.

For detailed guidance on the regulatory framework, see our post on PDPL vs GDPR for annotation vendors and the Saudi Arabia Vision 2030 LLM push for the broader market context shaping demand for Najdi-specific annotation.

Related Reading

• Gulf (Khaleeji) Arabic Sentiment Analysis: What Models Get Wrong
• End-to-End Arabic Data Labeling: Project Case Study
• Why Global AI Teams Source Data Annotation From Saudi Arabia
• Arabic Data Labeling Service

Frequently Asked Questions