Dataset Card for Yor-Sarc

Yor-Sarc is a gold-standard dataset for sarcasm detection in Yorùbá, a tonal and morphologically rich low-resource African language spoken by over 50 million people. The dataset was created to address the scarcity of high-quality annotated resources for figurative language understanding in African NLP.

It contains 436 manually annotated Yorùbá text instances labeled for binary sarcasm classification.

---

Dataset Details

Dataset Description

Yor-Sarc aims to accelerate research in Yorùbá NLP and broader low-resource African language modeling by providing a carefully curated benchmark dataset for sarcasm detection.

Unlike coarse-grained sentiment datasets, Yor-Sarc focuses on fine-grained figurative language understanding, specifically sarcasm — a pragmatically complex phenomenon that depends on tone, context, and cultural cues.

Each instance:

• Is written in Standardized Yorùbá orthography (with diacritics)
• Is annotated with a binary label such that we could eventually have:
  • 0 = non-sarcastic
  • 1 = sarcastic
• Was independently labeled by three native Yorùbá speakers
• Preserves annotator agreement for uncertainty-aware modelling

The dataset achieves substantial inter-annotator agreement:

• Fleiss’ κ = 0.7660
• Pairwise Cohen’s κ = 0.6732–0.8743
• 83.3% unanimous agreement

Yor-Sarc provides a foundational benchmark for evaluating machine learning, deep learning, and large language model (LLM) approaches for sarcasm detection in low-resource settings.

---

• Curated by: Toheeb Aduramomi Jimoh, Tabea De Wille, Nikola S. Nikolov
• Language(s): Yorùbá (yo)
• License: CC-BY-4.0

---

Dataset Sources

• Paper (arXiv): https://arxiv.org/abs/2602.18964
• ePrint: 2602.18964

---

Uses

Direct Use

Yor-Sarc is intended for research in:

• Sarcasm detection
• Figurative language understanding
• Sentiment and emotion analysis
• Low-resource NLP
• African language modelling
• Cross-lingual transfer learning
• Uncertainty-aware and disagreement-aware modeling

It is suitable for benchmarking:

• Machine learning models
• Deep learning architectures
• Transformer-based models
• Large Language Models (LLMs)

---

Out-of-Scope Use

This dataset should not be used for:

• Surveillance or profiling

---

Dataset Structure

Total instances: 436

Each record contains:

• text: Yorùbá sentence (with diacritics)
• label: Majority-vote binary sarcasm label
• (Optional) Annotator vote proportions for soft-label modeling

Agreement distribution:

• 83.26% unanimous agreement (3–0)
• 16.74% majority agreement (2–1)

---

Dataset Creation

Curation Rationale

Sarcasm detection research is heavily dominated by English and other high-resource languages. Yorùbá lacks publicly available gold-standard sarcasm datasets despite its linguistic richness and widespread usage.

Yor-Sarc was created to:

• Provide a culturally grounded benchmark
• Support African NLP research
• Encourage figurative language modeling in tonal languages
• Enable robust evaluation of sarcasm detection systems in low-resource settings

---

Source Data

Data were collected from:

• Yorùbá news platforms
• Social media platforms (X, Facebook, Instagram, YouTube captions)
• Crowdsourced contributions

All texts were cleaned and standardized before annotation.

---

Data Collection and Processing

• Web scraping, Manual collection and filtering
• Standardized orthographic normalization
• Three independent native-speaker annotations
• Majority voting for final labels
• Agreement metrics computed using Cohen’s κ and Fleiss’ κ

---

Annotations

Annotation Process

• Three native Yorùbá annotators
• Binary sarcasm labeling
• Iteratively refined guidelines
• Independent annotation
• Statistical agreement validation

Who are the annotators?

Native Yorùbá speakers with linguistic proficiency and familiarity with dialectal variation. Annotators were independent and compensated fairly.

---

Personal and Sensitive Information

The dataset contains publicly available and voluntarily contributed texts. Personally identifiable information has been removed where necessary. No sensitive metadata is included.

---

Bias, Risks, and Limitations

• Relatively small dataset size (436 instances)
• Non-exhaustive domain coverage
• Cultural understanding influences sarcasm interpretation

Citation

If you use Yor-Sarc, please cite:

BibTeX:

@misc{jimoh2026yorsarcgoldstandarddatasetsarcasm,
      title={Yor-Sarc: A gold-standard dataset for sarcasm detection in a low-resource African language}, 
      author={Toheeb Aduramomi Jimoh and Tabea De Wille and Nikola S. Nikolov},
      year={2026},
      eprint={2602.18964},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.18964}
}