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).