cs.CLNov 18, 2026

What Majority Labels Hide: Annotator Disagreement and Model Errors in Multi-Dialect Arabic Sexism Detection

Authors: Alaa Bazaid, Arkaitz Zubiaga

Organizations: School of Electronic Engineering and Computer Science, Queen Mary University of London, Mile End Road, London, E1 4NS, United Kingdom

Abstract

Judgements of sexism are subjective and may vary across communities, yet sexism detection datasets usually reduce annotator judgements to a single majority label, obscuring disagreement and how models behave on contested cases. We introduce MuDAS, a dataset of YouTube comments in Saudi, Egyptian and Moroccan Arabic, each labelled by three native speakers of its dialect, with the individual annotations preserved. We use it to evaluate Arabic encoders fine-tuned with hard-label, soft-label and perspectivist objectives, GPT-5.6 Terra, and four Arabic-centric large language models. Model errors are more frequent on comments with annotator disagreement. Despite relatively close Macro-F1 scores, GPT-5.6 Terra and the fine-tuned encoders exhibit contrasting error patterns: GPT-5.6 Terra, even with the annotators' definition of sexism, over-classifies gender-related content such as religious advice on modesty, whereas the encoders miss implicit and prescriptive sexism. The Arabic-centric models also over-predict sexism and remain below the encoders. Disagreement-aware learning preserves Macro-F1, but soft labels improve alignment with annotators on contested comments at the expense of unanimous ones, and neither paradigm reliably identifies contested comments. Lower performance on Egyptian Arabic, clearest for the large language models, is not fully explained by the examined measures of test-set composition and dialectness. Majority-label scores can therefore hide contrasting error patterns that matter for content moderation.

Figures & tables

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Aug 2, 2026

ArabicDialectSafety: A Dialect-Aware Benchmark for Arabic Content Safety Classification

We present ArabicDialectSafety, a human-curated Arabic safety dataset of 25,071 prompts covering six Arabic varieties: Modern Standard Arabic, Syrian, Egyptian, Algerian, Palestinian, and Moroccan. The dataset is annotated with dialect labels and seven fine-grained harm categories. We introduce a dual-task evaluation framework for binary safe/unsafe detection and granular harm classification across dialects. Benchmarking seven supervised and generative models, we find that fine-tuned MARBERTv2 achieves the strongest performance, with Macro-F1 scores of 0.95 for binary classification and 0.90 for granular classification, substantially outperforming prompted frontier LLMs, including Arabic-specialized models. Our analyses show that dialect conditioning is most effective when integrated at the representation level, while significant performance gaps remain for low-resource Maghrebi dialects. We further evaluate seven frontier LLMs as response generators on harmful dialectal Arabic prompts and observe unsafe generation rates below 5 percent across models. We release the dataset and code upon acceptance to support future research on dialect-aware Arabic safety evaluation. Warning: This paper contains examples of harmful and potentially offensive content included solely for research purposes.
Aug 30, 2026

Beyond Fluency: A Rubric-Based Benchmark for Evaluating Saudi Dialect and Cultural Competence in Large Language Models

Large language models are increasingly deployed in Arabic-speaking markets, yet standard benchmarks overwhelmingly reward Modern Standard Arabic (MSA) fluency while leaving dialectal and culturally grounded competence unmeasured. This gap is consequential: everyday Arabic is largely dialectal, and dialect encodes social meaning that MSA-centric evaluation cannot capture. We present a rubric-based benchmark for the Saudi dialect, comprising 31 expert-authored prompts spanning idiomatic, pragmatic, lexical, and culturally-embedded phenomena, each paired with an expert-established ground truth. Our methodology separates evaluation into a model-agnostic phase, in which atomic, MECE positive criteria are derived solely from the ground truth, and a model-specific phase, in which four state-of-the-art systems -- Claude Opus 5, Gemini 3.7, GPT-5.6, and Kimi K3 -- are scored against those criteria and penalised for errors they actively introduce. Across 124 model-prompt evaluations we catalogue 466 error instances under a nine-category taxonomy. The four systems cluster within a narrow macro-average band (42.7%-53.1%), with no model exceeding 55% and every model recording at least one negative-scoring prompt, confirming that Saudi dialectal competence remains broadly unsolved. Notably, Ambiguous Framing is the dominant failure mode (37.3% of errors) while outright Hallucination accounts for only 11.2%, indicating that models fail less by stating falsehoods than by distorting register and flattening pragmatic nuance. We further observe a consistency-versus-ceiling trade-off and model-distinctive error signatures. We release the full prompt set, ground truths, and scored rubrics to support reproducible dialectal evaluation.
Oct 13, 2026

GENDIS: Gender-Related Discourse and Stance in Multilingual YouTube Comments from India, Kenya, and Nigeria

Research on gender-related online content is often operationalized through harm-centered tasks that collapse non-harmful gender-related and unrelated content, while existing datasets remain concentrated in high-resource languages and Global North settings. To bridge these gaps, we introduce GENDIS, a dataset of 9,600 multilingual, human-annotated YouTube comments from India, Kenya, and Nigeria (3,200 per country), annotated using a two-level framework that first identifies whether gender is referenced and then classifies gender-related content as negative, positive, or neutral. This separation provides a denominator for gender-related discourse independent of harm, enabling both prevalence estimation and characterization of its evaluative range. In the uniform random stratum, 23.8-35.0% of comments are gender-related, and among those with a resolved stance, 82% are neutral or positive. Across models, gender-reference detection is substantially easier than stance classification, especially for positive and neutral speech. Lightly supervised classifiers match frontier zero-shot large language models at substantially lower cost and outperform generic toxicity detection, while cross-country transfer remains limited. GENDIS provides a locally grounded resource for measuring when gender is invoked online and how that reference is framed.