Arabic-NLP-2026/context-aware-arabic-to-english-model-with-register

Context-Aware Arabic Dialect Translation Dataset

This repository contains the dataset and code for the paper "Context-Aware Dialectal Arabic Machine Translation with Interactive Region and Register Selection" (Anonymous Submission).

Contents

• —`context_aware_en_ar_v2.ipynb`: The main Google Colab notebook used for training and evaluation.
• —`balanced_dataset_ready.csv`: The full augmented dataset (57,600 sentence pairs) produced by our RBDA pipeline.
• —`train_dataset.csv`: The strict training split (95%).
• —`test_dataset.csv`: The unseen test split (5%) used for the results reported in the paper.
• —`code/`: Directory containing the training scripts and augmentation logic used to reproduce our results.
• —train_model_optimized.py: The main training loop for fine-tuning mT5.
• —build_dataset.py: The RBDA pipeline code.
• —requirements.txt: Python dependencies.

Dataset Structure

The columns in the CSV files are:

• —input: The source English text with control tags (e.g., [Egyptian] [Medical] I have a headache).
• —target: The target Arabic translation in the specific dialect.
• —region: The dialect label (Egyptian, Levantine, Gulf, etc.).
• —context: The social context (Medical, Travel, etc.).
• —style: The register (Formal/Informal).