TerjamaBench: A Culturally Specific Dataset for Evaluating Translation Models for Moroccan Darija

Moroccan Darija, the widely spoken dialect of Arabic in Morocco, is rich in cultural expressions, regional variations, and multilingual influences. Despite its prevalence, there is a lack of robust, culturally relevant datasets for evaluating models on Moroccan Darija, particularly for translation tasks. To address this gap, we introduce TerjamaBench, a dataset specifically designed to evaluate the performance of translation models on Moroccan Darija across various cultural and linguistic contexts. The benchmark includes entries in Darija written in both the Latin alphabet (Arabizi) and Arabic script, along with their corresponding English translations.

Dataset Curation

This dataset has been manually annotated and reviewed by AtlasIA community, and is intended to support research and development in AI technologies that understand and process Moroccan Darija.

Dataset Composition

The benchmark contains 850 entries, carefully curated and categorized to cover a diverse range of linguistic and cultural scenarios.

Note: Some Darija samples are written in diverse Arabizi forms or might be mapped to different English translations. Hence, beware of duplicates if using only Darija values.

Below is the breakdown of categories (topics):

Topic	Description	Number of samples
Common Phrases	Everyday expressions like greetings and common sayings.	136
Named Entities	Sentences with proper nouns, place names, cities, etc.	53
Numeric and Date Expressions	Sentences containing numbers, dates, or time expressions.	62
Educational	Sentences from domains like medical, legal, or scientific contexts.	73
Mixed Language Content	Sentences combining Darija with MSA, French, or English.	50
Idioms	Proverbs and sayings unique to Moroccan culture.	51
Humor	Jokes, puns, or humorous expressions.	50
Religion	Sentences containing religious terms or expressions.	66
Single Words	Isolated words to test basic translation capabilities.	163
Long Sentences	Sentences designed to test coherence in lengthy translations.	50
Incorrect Spellings	Sentences with slight spelling errors to evaluate model robustness.	50
Dialectal Variations	Sentences from different Moroccan regions (northern, eastern, southern).	46

Key Features

• Cultural and Linguistic Authenticity: Reflects real-world Moroccan usage and incorporates a mix of everyday expressions, technical terms, and culturally rich content such as proverbs and humor.
• Manually Curated: All entries are carefully written, translated, and reviewed by native Moroccan speakers to ensure accuracy and relevance.
• Extensive Scope: The dataset captures linguistic diversity with standard Darija as well as regional dialects, and addresses practical challenges like mixed language use and spelling variations.

Usage

This benchmark can be used to:

• Evaluate machine translation systems.
• Test multilingual language models for Moroccan Darija understanding.
• Develop culturally sensitive AI systems.

To use the dataset, load it from Hugging Face Hub and refer to the accompanying examples for integration into your workflows.

Learn more

Read our article to learn more about the dataset and the evaluation of SOTA models on the benchmark. Find it here.

Citation

If you use this dataset in your research or applications, please cite:

@dataset{TerjamaBench,
  title={TerjamaBench: A Culturally Specific Dataset for Evaluating Translation Models for Moroccan Darija},
  author={AtlasIA},
  year={2024},
  url={https://huggingface.co/datasets/atlasia/TerjamaBench/}
}

Acknowledgments

We thank the members of AtlasIA for their dedication and efforts in creating this benchmark.

Special recognition goes to the contributors: Aissam Outchakoucht, Chaymae Rami, Mahmoud Bidry, Zaid Chiech, Imane Momayiz, Abdelaziz Bounhar, Abir Arsalane, Abdeljalil ElMajjodi, Aymane ElFirdoussi, Nouamane Tazi, Salah-Eddine Iguiliz, Hamza Essamaali, Ihssane Nedjaoui, Yousef Khoubrane, Khaoula Alaoui, Salah-Eddine Alabouch, Adnan Anouzla, Bilal El Hammouchi, Anas Amchaar, Taha Boukhari, Mustapha Ajeghrir, Ikhlas Elhamly, Fouad Aurag, Omar Choukrani.