---

About

NumidiaNLP is a research initiative dedicated to Natural Language Processing for Maghrebi Arabic and North African digital language.

Our research focuses on the linguistic diversity observed across Algeria, Morocco, and Tunisia, including dialectal Arabic, non-standard orthography, Arabizi, and multilingual code-switching.

The broader objective is to develop reusable language models, datasets, and evaluation resources for underrepresented Maghrebi language varieties.

---

Research Focus

• Maghrebi Arabic language modeling
• Regional continued pretraining
• Algerian, Moroccan, and Tunisian Arabic
• Arabizi and Latin-script Maghrebi Arabic
• Arabic–French and multilingual code-switching
• Hate and offensive language detection
• Sentiment analysis
• Dialect identification
• Corpus construction and annotation
• Robust and reproducible NLP evaluation

---

🌍 MagharibiBERT

A Regionally Adapted Language Model for Maghrebi Arabic

MagharibiBERT is a regional language model developed to provide reusable contextual representations for heterogeneous Maghrebi Arabic.

It is adapted through large-scale continued pretraining over diverse Maghrebi text representing linguistic variation from Algeria, Morocco, and Tunisia.

The training resources preserve important characteristics of real-world Maghrebi communication, including dialectal variation, non-standard orthography, Arabizi, and multilingual code-switching.

The central research question is:

Can regional continued pretraining provide transferable language representations across Maghrebi countries and NLP tasks?

🤗 Model:
https://huggingface.co/NumidiaNLP/MagharibiBERT

---

Regional Perspective

Maghrebi Arabic is not treated as a single homogeneous dialect.

Instead, we consider it a linguistically diverse regional continuum with substantial variation across countries, communities, scripts, platforms, and communication contexts.

Our research therefore investigates both:

• the value of shared regional representations, and
• the complementary strengths of country-specific models.

---

Evaluation Principles

Our experimental methodology emphasizes:

• fixed train/validation/test partitions;
• multiple random seeds;
• validation-based checkpoint selection;
• Macro-F1 for imbalanced classification tasks;
• comparisons with general Arabic and regional language models;
• statistical analysis where appropriate;
• reporting of computational requirements;
• transparent reporting of limitations and negative results.

The objective is not merely to obtain the highest score on one dataset, but to study cross-country robustness, stability, and transferability.

---

Resources

---

Reproducible Research

Where publication, licensing, and ethical requirements permit, NumidiaNLP aims to release:

• pretrained models;
• evaluation code;
• benchmark configurations;
• dataset documentation;
• reproducible experimental protocols.

Each released resource is accompanied by documentation describing its provenance, intended use, evaluation setting, and limitations.

---