JosueG/french-adja-parallel-corpus

French-Adja Parallel Corpus

The first publicly available parallel text corpus for Adja machine translation, targeting an under-resourced Gbe language spoken by approximately 1,000,000 people in Benin and Togo.

Dataset Description

This dataset contains 10,000 French-Adja sentence pairs created through a six-month collaborative translation effort with native Adja speakers in the Couffo region of Benin. It is designed to serve as a foundation for machine translation research and other NLP tasks for the Adja language.

About Adja

Adja (ISO 639-3: ajg) is a Gbe language of the Niger-Congo family, closely related to Fon, Ewe, and Gen. It is spoken by approximately 1 million people, primarily in the Couffo and Mono departments of southern Benin and in southeastern Togo. Despite its significant speaker population, Adja had no publicly available text-based NLP resources prior to this work --- no parallel corpora, no machine translation systems, and no labeled computational datasets. Concurrent work by Justin et al. (2025) introduced Eyaa-Tom, a multi-language dataset for Togolese languages that includes Adja among its targets, but the publicly released Adja data consists only of a small amount of audio — no French-Adja parallel text for MT has been made available.

Languages

	Language	ISO 639	Script
fr	French	fra	Latin
adj	Adja	ajg	Latin (with diacritics: ɔ, ɛ, ɖ, ŋ, tonal marks)

Dataset Creation

Translation Process

1. Source sentences: 10,000 French sentences selected through uniform random sampling from Tatoeba, a collaborative platform of community-contributed translations
2. Translation team: 5 native Adja speakers from the Couffo region of Benin:
   • 2 translators (1 government-accredited Adja language instructor + 1 experienced fluent speaker)
   • 3 transcribers (all native speakers)
3. Process: French sentences were read aloud, discussed for comprehension, then translated orally into Adja. Transcribers recorded the spoken Adja in writing. Translations were then typed and formatted for computational use
4. Duration: 6 months of collaborative work
5. Quality control: A dedicated cleaning pipeline was applied:
   • Unicode normalization (NFKC)
   • Spacing normalization
   • Non-standard character replacement
   • Bidirectional punctuation consistency checks
   • Quotation mark matching

Why Oral Translation?

Adja is primarily a spoken language. Having translators produce Adja translations orally before transcription preserves natural spoken Adja and avoids the artificiality that can arise from written-first translation in a language with limited written tradition.

Dataset Structure

Data Fields

Field	Type	Description
fr	string	French source sentence
adj	string	Adja translation

Data Splits

Split	Count	Percentage
train	8,000	80%
validation	1,000	10%
test	1,000	10%

Splits were created through uniform random sampling (seed 42).

Examples

French	Adja
Je sue tous les jours.	ŋ kɔ nɔ ade tɛgbɛ ɛ.
Je pense que tu devrais voir ça.	ŋ bumɔ́ wɔ a kpɔɛ alo wɔ a nya.
Tu te perds.	E búbu ɔ deki.
Cesse de rêver et ouvre les yeux.	Mi edrɔ kukú ahùn ŋkuvi wo.
Le gouvernement a reçu son autorité de l'empereur.	Eju tatɔ xɔ acɛ ega ɖuɖu tɔ.

Corpus Statistics

Metric	French	Adja
Total tokens	66,245	66,661
Vocabulary size	11,385	12,560
Type-Token Ratio	0.172	0.188
Hapax legomena	7,196 (63%)	8,318 (66%)
Mean sentence length	6.62 words	6.67 words
Median sentence length	6.0	6.0
Sentence length std	2.92	3.18
Min -- Max length	1 -- 63	1 -- 68

Adja exhibits higher lexical diversity (TTR 0.188 vs 0.172), reflecting the morphological richness characteristic of Gbe languages.

Baseline Results

We fine-tuned three models on this corpus and report mean results over 5 random seeds on the random test split:

Model	Direction	BLEU	chrF++
NLLB-600M	FR → ADJ	4.5 ± 0.2	26.1 ± 0.3
NLLB-600M	ADJ → FR	11.8 ± 0.6	30.3 ± 0.5
mBART-50	FR → ADJ	3.4 ± 0.3	22.6 ± 0.5
mBART-50	ADJ → FR	8.7 ± 0.1	26.1 ± 0.3
ByT5-base	FR → ADJ	4.3 ± 0.5	26.1 ± 0.4
ByT5-base	ADJ → FR	11.5 ± 0.9	31.2 ± 0.9

All models used Adafactor (lr=1e-4), batch size 16, and early stopping on validation chrF (patience 10).

Usage

from datasets import load_dataset

dataset = load_dataset("JosueG/french-adja-parallel-corpus")

# Access a sentence pair
print(dataset["train"][0])
# {'fr': 'Je sue tous les jours.', 'adj': 'ŋ kɔ nɔ ade tɛgbɛ ɛ.'}

# Fine-tune a translation model
for example in dataset["train"]:
    src = example["fr"]
    tgt = example["adj"]

Ethical Considerations

• This corpus was created in direct collaboration with native Adja speakers in Benin. The translation team was compensated for their work
• The dataset is released under a non-commercial license (CC BY-NC-SA 4.0) to protect against exploitative use while enabling academic research
• Adja is primarily a spoken language; this written corpus does not capture the full richness of spoken Adja, including tonal variation and dialectal differences across communities
• The Tatoeba source sentences reflect global French usage and may not represent the specific variety of French spoken in Benin and Togo

Citation

If you use this dataset, please cite:

@inproceedings{godeme2026french-adja,
    title     = {A 10,000-Sentence French-Adja Parallel Corpus for Machine Translation},
    author    = {Godeme, Josue and Coto-Solano, Rolando},
    booktitle = {Proceedings of the 2026 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2026)},
    year      = {2026},
    note      = {To appear}
}

Acknowledgements

We are deeply grateful to the Adja-speaking community members in the Couffo region of Benin who dedicated their time and expertise to translating and transcribing this corpus over a six-month period. Their commitment to documenting and advancing their language made this work possible.