Dataset Card for ArDQA

ArDQA is a cross-dialect Arabic QA benchmark spanning three domains. Each domain provides parallel QA triples {context, question, answer} across five Arabic varieties: MSA, Egyptian, Gulf, Levantine, Maghrebi. The benchmark contains 8,150 QA triples overall and is designed for evaluation of cross-dialectal transfer in Arabic extractive QA.

Dataset Details

Composition

• ArDQA-SQuAD: Curated from Arabic-SQuAD v2.0, then translated by native speakers into four dialects with manual span annotation to preserve one-to-one alignment.
• ArDQA-Vlogs: Colloquial lifestyle vlog transcripts --> QA construction --> dialect translations --> manual span annotation.
• ArDQA-Narratives: Cultural narratives and folklore from online videos, following the same pipeline as Vlogs, with longer, descriptive answers.

Quality control

Native speakers translated independently in every domain, cross-checked each other, and an expert adjudicated disagreements. Span consistency was validated (using answer-to-context length ratios) to maintain strict alignment across dialects.

Paper

Under Review

Direct Use

• Evaluation of zero-shot and few-shot cross-dialectal transfer in Arabic QA.
• Analysis of dialectal robustness for Arabic extractive QA models.
• Benchmarking domain sensitivity across SQuAD-like, vlog, and narrative content.

Dataset Structure

Format

ArDQA follows SQuAD v2.0 JSON:

root
├── data: [
│   ├── {
│   │   ├── title: string
│   │   └── paragraphs: [
│   │       ├── {
│   │       │   ├── context: string
│   │       │   └── qas: [
│   │       │       ├── {
│   │       │       │   id: string
│   │       │       │   question: string
│   │       │       │   is_impossible: boolean
│   │       │       │   answers: [
│   │       │       │       ├── { text: string, answer_start: int }
│   │       │       │       └── ...
│   │       │       └── ...
│   │       └── ...
│   └── ...
└── (optional) version: string

Splits

Each ArDQA domain is divided into development and test splits to enable zero-shot evaluation (train on MSA or other sources, then evaluate on dialects without target-dialect fine-tuning).

Counts per domain and split

fold	ArDQA-SQuAD (# parallel / # total)	ArDQA-Vlogs (# parallel / # total)	ArDQA-Narratives (# parallel / # total)
dev	131 / 655	171 / 855	160 / 800
test	368 / 1,840	436 / 2,180	364 / 1,820

• # parallel = Number of {context, question, answer} triples aligned across all five Arabic varieties.
• # total = # parallel × 5 dialects (MSA, Egyptian, Gulf, Levantine, Maghrebi).
• Totals across all domains: dev = 2,310, test = 5,840, overall = 8,150 QA triples.

Source Data

Original texts come from Arabic-SQuAD v2.0 and public online video transcripts (vlogs, narratives). QA items and dialect translations were produced by native-speaker annotators.

Annotations

Annotation process

• Native speakers independently translate and annotate spans.
• Cross-review and expert adjudication.
• Consistency checks (e.g., answer length vs. context, span alignment across dialects).

Experiment (brief)

We evaluate zero-shot cross-dialectal transfer by training only on MSA (Arabic-SQuAD v2.0) and testing zero-shot on dialectal data.

• Models: AraELECTRA-MSA-QA, CAMeLBERT-MSA-QA, AraBERT-MSA-QA.
• Data: ArDQA dev/test across three domains (SQuAD, Vlogs, Narratives) and five varieties: MSA, Egyptian (EGY), Gulf (GLF), Levantine (LEV), Maghrebi (MGR).

Model References

• AraELECTRA-MSA-QA. Hugging Face model card. https://huggingface.co/ZeyadAhmed/AraElectra-Arabic-SQuADv2-QA
• CAMeLBERT-MSA (bert-base-arabic-camelbert-msa). Hugging Face model card. https://huggingface.co/CAMeL-Lab/bert-base-arabic-camelbert-msa
• AraBERT-MSA-QA (bert-large-arabertv02). Hugging Face model card. https://huggingface.co/aubmindlab/bert-large-arabertv02

Evaluation Metrics

• EM (Exact Match): 1 if the predicted span matches the gold answer exactly; else 0.
• F1: token-level harmonic mean of precision and recall between predicted and gold spans (rewards partial overlap).

Reference Results (Zero-Shot Cross-Dialectal Transfer)

ArDQA-SQuAD (F1 / EM)

Model	EGY	GLF	LEV	MGR	MSA
AraELECTRA-MSA-QA	71.66 / 59.51	73.76 / 60.87	66.72 / 50.54	66.35 / 53.80	76.19 / 61.96
CAMeLBERT-MSA-QA	53.98 / 28.04	54.91 / 26.68	51.49 / 25.86	46.90 / 23.96	60.27 / 29.13
AraBERT-MSA-QA	12.53 / 4.04	11.01 / 3.74	12.01 / 3.88	12.06 / 3.54	11.80 / 3.74

ArDQA-Vlogs (F1 / EM)

Model	EGY	GLF	LEV	MGR	MSA
AraELECTRA-MSA-QA	63.90 / 37.93	64.47 / 41.74	63.00 / 40.37	57.11 / 31.19	67.01 / 42.20
CAMeLBERT-MSA-QA	40.66 / 15.09	39.12 / 14.63	37.50 / 14.17	29.69 / 10.04	46.66 / 16.01
AraBERT-MSA-QA	13.08 / 4.03	11.18 / 4.03	12.03 / 4.45	12.49 / 4.03	11.53 / 4.68

ArDQA-Narratives (F1 / EM)

Model	EGY	GLF	LEV	MGR	MSA
AraELECTRA-MSA-QA	35.75 / 11.26	40.80 / 14.20	38.31 / 12.98	31.70 / 6.87	43.83 / 14.05
CAMeLBERT-MSA-QA	22.33 / 5.82	25.53 / 8.02	20.74 / 6.56	23.82 / 7.47	25.13 / 9.68
AraBERT-MSA-QA	16.20 / 4.01	16.82 / 4.27	15.41 / 4.01	18.72 / 4.27	18.07 / 4.01

Citation

If you use ArDQA, please cite the dataset

Dataset

BibTeX

@dataset{ardqa_dataset_2025,
  title   = {ArDQA: Cross-Dialect Arabic QA Benchmark},
  authot =  {Althobaiti, Maha Jarallah}
  year    = {2025},
  note    = {Hugging Face dataset, CC BY-NC-ND 4.0},
  url     = {https://huggingface.co/datasets/MahaJar/ArDQA}
}