sba-Fr: Ngambay–French Parallel Corpus

sba-Fr is the first public parallel corpus between Ngambay (ISO 639-3 sba) and French. Ngambay is a Sara (Central Sudanic) language spoken mainly in southern Chad (Logone Occidental, Logone Oriental, Tandjilé, Mayo-Kebbi) and northern Cameroon, and is used as a lingua franca across southwestern Chad.

The corpus contains 33,073 human-produced sentence pairs, plus an optional set of 5,000 machine-generated (synthetic) pairs. It was released with the paper Ngambay-French Neural Machine Translation (sba-Fr) (NLP4TIA workshop at RANLP 2023), where M2M100, ByT5 and mT5 were fine-tuned on it.

• Paper: ACL Anthology 2023.nlp4tia-1.6 · arXiv:2308.13497
• Code & raw files: GitHub – Toadoum/Ngambay-French-Neural-Machine-Translation-sba_fr_v1-
• Curated by: Sakayo Toadoum Sari, Angela Fan, Lema Logamou Seknewna
• Point of contact: Sakayo Toadoum Sari

Quick start

from datasets import load_dataset

ds = load_dataset("Toadoum/ngambay-french-sba-fr")                       # config "original" (default)
ds_syn = load_dataset("Toadoum/ngambay-french-sba-fr", "original_plus_synthetic")

print(ds["train"][0])
# {'id': 'original-train-00000', 'sba': '18Yeḛ gə́ iya dɔ kḛji loo ...', 'fr': '18 Celui qui dissimule la haine ...', 'source': 'bible'}

Configurations

Config	train	validation	test	Contents
original (default)	21,166	6,615	5,292	Human-produced pairs only. Exact splits used in the paper (Table 1).
original_plus_synthetic	24,366	7,615	6,092	The 33,073 original pairs re-shuffled together with 5,000 synthetic pairs. Exact splits used in the paper (Table 2).

⚠️ The two configs use different splits. About 64% of the original test set (3,397 / 5,292 pairs) appears in the original_plus_synthetic train split. Never train on one config and evaluate on the other.

Data fields

Field	Type	Description
id	string	Stable row identifier (<config>-<split>-<index>).
sba	string	Ngambay text.
fr	string	French text.
source	string	bible, dictionary or synthetic (see below).

Composition by source

Source	original	original_plus_synthetic	Description
bible	31,897	31,897	Verse-aligned pairs scraped from YouVersion (Ngambay Bible ↔ French Louis Segond).
dictionary	1,176	1,176	Everyday sentences created manually from the Sara Bagirmi Languages Project Ngambay–French dictionary (5th ed., 2015).
synthetic	–	5,000	French news sentences (Leipzig Corpora Collection, fra_news_2022_100K) translated into Ngambay by the paper's fine-tuned M2M100 model.

Dataset creation

Bible data. Parallel verses were scraped from YouVersion with R (Dataset/bible_scrapping.Rmd in the GitHub repo). The Ngambay translation does not cover every verse of the French version, and the raw scrape contained errors, incomplete translations and duplicates. Native speakers and linguists, including members of the association translating the Bible from French into Ngambay in Chad, reviewed the data; inconsistent and incomplete pairs were dropped.

Dictionary data. 1,176 short-to-medium everyday sentences were entered manually (via a Google Form) from the Ngambay–French dictionary of the Sara Bagirmi Languages Project.

Synthetic data. Following forward translation / self-training (Sennrich et al., 2016; He et al., 2020), the best model trained on the original data (M2M100) translated French monolingual news sentences into Ngambay. The Ngambay side of these pairs is model output, not human translation, and has not been verified by speakers.

Split. The original corpus was split roughly 64/20/16 into train/validation/test.

Differences from the GitHub release

This Hub version is the same data, with the following changes:

1. Language key renamed sw → sba. In the GitHub JSON files the Ngambay side is stored under sw (Swahili) because Swahili was used as the proxy language code when fine-tuning M2M100/mT5/ByT5. The text itself is Ngambay.
2. Leipzig sentence IDs removed from the French side of synthetic rows (e.g. "1895\tAlep-Sana/ …" → "Alep-Sana/ …").
3. source column added, derived from the data: synthetic = not present in the original corpus; bible = French side starts with a verse number; dictionary = the rest (exactly 1,176 rows, matching the paper).
4. Not included: sba_fr_JSON/*_synthetic.json in the GitHub repo, which adds 1,574 non-Bible pairs whose provenance is not documented in the paper. The splits from Dataset/*_sy.json (the ones matching the paper's Table 2) are used instead.

The text is otherwise unmodified, so results remain comparable with the paper.

Known issues and limitations

• Domain bias. ~96% of the human-produced data is biblical. Expect archaic register, religious vocabulary, and weak coverage of modern topics.
• Verse numbers kept in the text. Bible rows begin with the verse number on both sides ("18 Celui qui…" / "18Yeḛ gə́…"; no space on the Ngambay side). Strip them with re.sub(r'^\d+\s*', '', text) for most uses, and do so consistently for both training and evaluation.
• Partial alignments. Some Ngambay verses are shorter than their French counterpart because verse boundaries differ between translations. About 1,790 training pairs have a Ngambay/French word ratio below 0.3.
• Orthographic inconsistency. Dictionary rows use a tone-marked orthography (e.g. gɨ́, ɗ, ā), while Bible rows use the church orthography (e.g. gə́, d’, without systematic tone marking). Normalise to NFC and be aware that the same word may appear in both spellings.
• Duplicates and overlap. A few exact duplicates exist within splits (≤55 per split) and a small number of pairs/source sentences occur in both train and test (13 identical pairs, 130 shared Ngambay sentences in original).
• Synthetic rows in evaluation splits. In original_plus_synthetic, validation and test also contain synthetic rows (967 and 781). Scores on these splits partly measure agreement with an earlier model's output. For a clean evaluation, filter with ds["test"].filter(lambda x: x["source"] != "synthetic").

Licensing

The data is released for research and non-commercial use.

• The code and curation in the GitHub repository are MIT-licensed, but that licence does not cover third-party text.
• French Bible text (Louis Segond 1910) is in the public domain.
• Ngambay Bible text was obtained from YouVersion; copyright remains with its publisher. Users must check the publisher's terms before any commercial use or redistribution.
• Synthetic rows derive from the Leipzig Corpora Collection; see its terms of use.
• Dictionary rows were created by the authors from the Sara Bagirmi Languages Project dictionary.

Citation

@inproceedings{toadoum-sari-etal-2023-ngambay,
    title     = "{N}gambay-{F}rench Neural Machine Translation (sba-{F}r)",
    author    = "Toadoum Sari, Sakayo and Fan, Angela and Seknewna, Lema Logamou",
    booktitle = "Proceedings of the First Workshop on NLP Tools and Resources for Translation and Interpreting Applications",
    month     = sep,
    year      = "2023",
    address   = "Varna, Bulgaria",
    publisher = "INCOMA Ltd., Shoumen, Bulgaria",
    url       = "https://aclanthology.org/2023.nlp4tia-1.6",
    pages     = "39--47",
}

Acknowledgements

This work was carried out at AIMS / AMMI (African Master's in Machine Intelligence) and supported by a Google Cloud Platform grant, as part of the Masakhane community effort for African-language MT. We thank the Ngambay speakers and Bible translators who reviewed the data.