| | --- |
| | name: Zarma Noisy Dataset |
| | language: |
| | - dje |
| | version: 1.0.0 |
| | license: cc-by-sa-4.0 |
| | source: Derived from monolingual Zarma dataset |
| | task_categories: |
| | - text-classification |
| | - text-generation |
| | - fill-mask |
| | - question-answering |
| | size_categories: |
| | - 100K<n<1M |
| | --- |
| | |
| | # Zarma Noisy Dataset |
| |
|
| | ## Overview |
| |
|
| | The Zarma Noisy Dataset is a collection of Zarma sentences with artificially introduced noise to simulate human-like errors. This dataset is designed for tasks such as grammatical error correction (GEC), text denoising, and robustness testing in natural language processing (NLP) for low-resource languages like Zarma. It is derived from a clean monolingual Zarma dataset (`monolingual_zarma.jsonl`) by applying various types of noise, including character-level and word-level modifications. |
| |
|
| | ## Dataset Structure |
| |
|
| | The dataset is stored in a JSONL file (`noisy/zarma_noisy_dataset.jsonl`) where each line is a JSON object with the following fields: |
| |
|
| | * **`original`**: The raw input sentence as it appears in the source dataset, preserving its exact form. |
| | * **`cleaned`**: A normalized version of the sentence (Unicode NFC normalization, extra spaces removed). |
| | * **`char_swap`**: The sentence with adjacent character swaps (e.g., "teh" → "the") within words to mimic typos. |
| | * **`random_char_insertion`**: The sentence with up to 2 random character insertions, preferring vowels near vowels for realism. |
| | * **`char_delete`**: The sentence with character deletions, avoiding critical positions (first/last in words). |
| | * **`char_substitute`**: The sentence with character substitutions, using similar-looking or keyboard-adjacent characters (e.g., 'a' → 's'). |
| | * **`word_masking`**: The sentence with words replaced by a `BLANK` token, preferring content words (length > 3). |
| | * **`word_swap`**: The sentence wit adjacent word swaps (e.g., "is it" → "it is"). |
| | |
| | ### Example Entry |
| | |
| | ```json |
| | { |
| | "original": "Yesu Kirisita Tuura Wema TUURA WEMA", |
| | "cleaned": "Yesu Kirisita Tuura Wema TUURA WEMA", |
| | "char_swap": "Yseu Kirisita Tuura Wema TUURA WEMA", |
| | "random_char_insertion": "Yesu Kirisita Tuura Wema TUURA aWEMA", |
| | "char_delete": "Yesu Kirista Tuura Wema TURA WEMA", |
| | "char_substitute": "Yesu Kirisita Tuura Wema TUURA WEMs", |
| | "word_masking": "Yesu Kirisita BLANK Wema TUURA WEMA", |
| | "word_swap": "Yesu Kirisita Wema Tuura TUURA WEMA" |
| | } |
| | ``` |
| | ## Citation |
| | |
| | If you use this dataset in your research, please cite the following paper: |
| | code |
| | Bibtex |
| | ```bibtex |
| | @misc{keita2025grammaticalerrorcorrectionlowresource, |
| | title={Grammatical Error Correction for Low-Resource Languages: The Case of Zarma}, |
| | author={Mamadou K. Keita and Christopher Homan and Marcos Zampieri and Adwoa Bremang and Habibatou Abdoulaye Alfari and Elysabhete Amadou Ibrahim and Dennis Owusu}, |
| | year={2025}, |
| | eprint={2410.15539}, |
| | archivePrefix={arXiv}, |
| | primaryClass={cs.CL}, |
| | url={https://arxiv.org/abs/2410.15539}, |
| | } |
| | ``` |
