IWAN/adab-arabic-politeness

ADAB – Arabic Dataset for Automated Politeness Benchmarking

📄 Paper: ADAB: Arabic Dataset for Automated Politeness Benchmarking — a Large-Scale Resource for Computational Sociopragmatics — LREC 2026

ADAB is a large-scale Arabic corpus for studying politeness computationally (computational sociopragmatics). Each text is labelled as Polite, Neutral, or Impolite, and is drawn from four online sources spanning social media, e-commerce, and company/app reviews.

Dataset summary

Split	Rows
Train	4,895
Validation	693
Total	5,588

Fields

Field	Description
Sentence	The Arabic text
Source	Origin of the text: Tweet, Companies, Youtube, or Shein
label	Politeness label: Polite, Neutral, or Impolite

Label distribution

Label	Train	Validation
Neutral	3,664	519
Polite	815	115
Impolite	416	59

Source distribution (train)

Source	Rows
Tweet	1,246
Companies	1,229
Youtube	1,213
Shein	1,207

Citation

If you use ADAB, please cite the LREC 2026 paper:

@inproceedings{alkhalifa2026adab,
  title     = {ADAB: Arabic Dataset for Automated Politeness Benchmarking --
               a Large-Scale Resource for Computational Sociopragmatics},
  author    = {Al-Khalifa, Hend and Ghezaiel, Nadia and Bounnit, Maria and
               Alhazmi, Hend Hamed and Alfear, Noof Abdullah and
               Alqifari, Reem Fahad and Almasoud, Ameera Masoud and
               Al-Ghamdi, Sharefah Ahmed},
  booktitle = {Proceedings of the 2026 Conference on Language Resources
               and Evaluation (LREC 2026)},
  year      = {2026},
  publisher = {European Language Resources Association (ELRA)},
  url       = {https://aclanthology.org/2026.lrec-main.244/}
}