Alexandria

Paper and Dataset

Alexandria: English to Local Arabic Dialect Conversations at Scale

Alexandria is a community-built parallel dataset of English and dialectal Arabic conversations across high social-impact domains, featuring human-translated and human-revised content designed for culturally grounded Arabic MT, natural code-switching, and inclusive gender directions (male and female), while capturing diverse scenarios with specific personas and roles.

Coverage
Spans healthcare, education, legal and financial services, logistics, tourism, and other public-facing domains.
Structure
Preserves multi-turn conversational context with turn-level English-dialect alignment, enriched with persona metadata including roles and gender.
Linguistic Focus
Includes subdialects, naturally occurring code-switching across several countries, and registers adapted to specific conversation scenarios.

Paper Summary

Highlights from the paper

Scale and Coverage

Alexandria comprises 34,488 conversations and 107,631 turns, spanning 13 Arab countries and 11 high-impact social domains. It features city-level grounding and rich multi-turn conversational context.

Metadata Depth

Alexandria’s conversations are enriched with metadata for city-level dialects, persona roles, and speaker-addressee gender, supporting both context- and gender-aware evaluation.

Revision Quality Signals

The Alexandria Dataset was developed in two human-centric phases. In the first phase, participants translated the source data from English into specific Arabic dialects. This was followed by a review and revision phase, where 31.6% of the turns were modified.

Gender and Code-Switching

Alexandria features metadata for city-level varieties and gender-specific interactions, specifically female-to-male, male-to-female, male-to-male, and female-to-female. To ensure stylistic accuracy, registers were adapted based on the scenario and persona, enabling a context-aware evaluation that reflects natural code-switching and social dynamics.

Evaluation Setup

The paper evaluates 24 Arabic-capable LLMs under turn-level, context-level, and conversation-level prompting with metadata-aware inputs, and reports spBLEU and chrF++ for automatic assessment.

Main Benchmark Findings

Results show a strong directional asymmetry: dialect-to-English is easier than English-to-dialect. Human evaluation reports high gender adherence, while dialectness remains a major challenge for many systems.

Dataset Stats

Coverage and scale

107,631

Total human-translated English-to-dialect turns

13

Arab countries

11

Domains

4

Regional groups: Levant, Gulf, Nile, and Maghreb

Swipe horizontally in the table to view all countries and totals.

Domain Levant Gulf Nile Maghreb
JO LB PS SY SA OM YE EG SD LY MA MR TN
Agriculture/Farming 8251,1401,7709311,162915529583163231570970481
Commerce/Transactions 7501,0041,5957491,020650579506201160445757401
Construction/Real Estate 8599951,7618611,161974696660225271574673485
Education/Academia 8161,1911,5138311,0171,079563549170220601863551
Energy/Resources 7861,0481,7159281,177937587625189243447719470
Everyday/Social 9671,2151,6977871,020888642604175210595824550
Healthcare/Medical 7271,2401,7287811,043895548487164253556948522
Legal/Financial 6931,0061,566757857753496539177174481642412
Logistics/Transport 8421,0201,5129501,234842629646189187593877515
Professional/Workplace 8451,2201,8109591,112866549645178253480709526
Tourism/Hospitality 7201,1611,5968841,004815608608190216567878460
Total 8,83012,24018,2639,41811,8079,6146,4266,4522,0212,4185,9098,8605,373

Qualitative Examples

Samples from Alexandria Dataset

Resources

Alexandria Resources

Available now

Paper

Read the full Alexandria research paper on ACL Anthology.

Open ACL Anthology
Available now

Data

Browse the Alexandria dataset release on Hugging Face, including splits and hosted dataset files.

Open Dataset on 🤗
Available now

Code

Access the Alexandria project repository on GitHub for evaluation code, updates, and related project materials.

Open GitHub

Authors

Alexandria team

Citation

How to cite Alexandria

If you use Alexandria or the associated methodologies, please cite the original paper using the following BibTeX entry.

@inproceedings{el-mekki-etal-2026-alexandria,
    title = "Alexandria: A Multi-Domain Dialectal {A}rabic Machine Translation Dataset for Culturally Inclusive and Linguistically Diverse {LLM}s",
    author = "EL Mekki, Abdellah  and
      Magdy, Samar M.  and
      Atou, Houdaifa  and
      AbuHweidi, Ruwa  and
      Qawasmeh, Baraah  and
      Nacar, Omer  and
      Al-hibiri, Thikra  and
      Saadie, Razan  and
      Alsayadi, Hamzah A.  and
      Hammouda, Nadia Ghezaiel  and
      Alkhazimi, Alshima Mohammed  and
      Hamod, Aya  and
      Al-Ghafri, Al-Yas Yaqoob  and
      El-Sayed, Wesam  and
      al Sharji, Asila Ismail  and
      Ballout, Mohamad  and
      Belfathi, Anas  and
      Ghaddar, Karim  and
      Sibaee, Serry  and
      Aoun, Alaa  and
      Aseri, Aeej Mohammed  and
      Abureesh, Lina  and
      Bashiti, Ahlam  and
      Yousef, Majdal  and
      Hafiz, Abdulaziz  and
      Mohamed, Yehdih  and
      Hamedtou, Emira  and
      Emehah, Brakehe  and
      Alhamouri, Rahaf  and
      Nafea, Youssef  and
      El Aatar, Aya  and
      Al-Dhabyani, Walid  and
      Hamed, Emhemed S.  and
      Shatnawi, Sara  and
      Alwajih, Fakhraddin  and
      Elkhidir, Khalid  and
      Alasmari, Ashwag  and
      Gerrio, Abdurrahman  and
      Alshahri, Omar Said  and
      Elmadany, AbdelRahim A.  and
      Berrada, Ismail  and
      Al-kathiri, Amir Azad Adli  and
      Zaraket, Fadi  and
      Jarrar, Mustafa  and
      EL Hadj, Yahya Mohamed  and
      Alhuzali, Hassan  and
      Abdul-Mageed, Muhammad",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.1503/",
    pages = "32567--32592",
    ISBN = "979-8-89176-390-6"
}