atlasia/ATAM

Overview

This dataset is a modified version of the Darija Open Dataset (DODa), tailored specifically for the purpose of learning a transliteration mapping from Arabizi Darija text to Arabic letters Darija.

The Arabizi-To-Arabic-Mapping (ATAM) Transliteration Dataset serves as a valuable resource for training models to accurately transliterate Arabizi Darija text into Arabic letters Darija, facilitating natural language processing tasks in the Moroccan dialect.

Key Features:

• Adapted Structure: The dataset has been meticulously adapted from the original DODa format to focus on the transliteration task, ensuring relevance and effectiveness in training transliteration models.

• Diverse Text Samples: It encompasses a diverse range of Arabizi Darija text samples, covering various linguistic nuances and expressions commonly found in informal communications.

• Annotated Transliterations: Each Arabizi Darija text entry is accompanied by its corresponding transliteration into Arabic letters Darija, enabling supervised learning for transliteration mapping.

N.B: We have a One-To-Many relationship as each word in the Arabic letter format can be associated (written) to many others in the Arabizi format.

Usage:

Researchers and developers can leverage this dataset to train and evaluate machine learning models aimed at automating the transliteration process from Arabizi Darija to Arabic letters Darija.

Acknowledgments:

We extend our gratitude to the creators and contributors of the Darija Open Dataset (DODa) for their pioneering work, which serves as the foundation for this transliteration dataset adaptation.

Contact:

For inquiries or contributions related to the ATAM Transliteration Dataset, feel free to reach out.

---

dataset_info: features: - name: darija dtype: string - name: darija_ar dtype: string - name: english dtype: string splits: - name: train num_bytes: 2705541 num_examples: 67186 download_size: 1793167 dataset_size: 2705541 configs: - config_name: default data_files: - split: train path: data/train-* language: - ar size_categories: - 10K<n<100K