glk360/fr-ewe-corpus

FR-EWE Parallel Corpus v0.0

A sentence-level French → Ewe (Éwé) parallel corpus focused on the civic rights domain. Derived entirely from authoritative public-domain sources and reviewed by a native Ewe speaker.

Dataset Description

Languages: French (fr) · Ewe / Éwé (ee, ISO 639-1) Domain: Civic rights — Universal Declaration of Human Rights (UDHR) Version: 0.0 (seed corpus — 60 sentence pairs) License: CC-BY 4.0 (source documents are public domain; attribution required for derivative works)

Why Ewe?

Ewe is spoken by approximately 7 million people across Togo, Ghana, and Benin. Despite this population, no production-grade French → Ewe NLP pipeline exists and publicly available parallel corpora are extremely sparse. This dataset is a first building block.

What's in v0.0

Alignment method breakdown:

All 60 pairs reviewed and confirmed by a native Ewe speaker (Westland-corp, Lomé, Togo).

Note on precision-matched pairs: 17 pairs where the French is a short clause and the Ewe is the full containing paragraph. Filter on confidence >= 0.80 to exclude these if you need symmetric-length pairs.

Schema

{
  "fr_text":    str,   # French source sentence
  "ewe_text":   str,   # Ewe translation
  "source":     str,   # Source document and article (e.g. "DUDH — Article 1")
  "domain":     str,   # "civique"
  "match_type": str,   # "exact" | "sub_sentence" | "fuzzy" | "precision"
  "confidence": float  # Alignment confidence (1.0 = exact)
}

Source Attribution

French and Ewe texts from the Unicode UDHR corpus (eric-muller/udhr), itself derived from the OHCHR official translations. The Universal Declaration of Human Rights is in the public domain.

Curation

This dataset was created by GLK 360 (Windsor Locks, CT, USA) in collaboration with Westland-corp (Lomé, Togo).

French sentences were extracted verbatim from the UDHR. Ewe counterparts were aligned automatically using the official UDHR Ewe XML via align_udhr.py (see ewe-localization-pipeline). All pairs were reviewed by Komlan Mensah, Ewe language specialist at Westland-corp.

Known Limitations

• —Domain: Civic rights only (v0.0). Health, agriculture, and general domains are in preparation.
• —Scale: 60 pairs. Suitable for evaluation and fine-tuning signal; not sufficient for training from scratch.
• —Alignment: 17 precision-matched pairs have asymmetric length. Filter on confidence >= 0.80 to exclude.
• —Script: Ewe uses the Latin script with extended characters (ƒ, ɖ, ɔ, ŋ, ʋ, ẽ). Ensure your tokenizer handles these correctly.

Related Resources

• —FR-EWE Pipeline (live demo): huggingface.co/spaces/glk360/fr-ewe-pipeline
• —Pipeline code: github.com/glk360/ewe-localization-pipeline
• —Models used: openai/whisper-medium · facebook/nllb-200-distilled-600M · facebook/mms-tts-ewe

Citation

@dataset{glk360_fr_ewe_corpus_2026,
  title     = {FR-EWE Parallel Corpus v0.0},
  author    = {Assignon, Kossivi and Mensah, Komlan},
  year      = {2026},
  publisher = {GLK 360},
  url       = {https://huggingface.co/datasets/glk360/fr-ewe-corpus},
  note      = {Civic rights domain. UDHR source. 60 sentence pairs.}
}

Contact

GLK 360 (Windsor Locks, CT, USA) — info@glk360.com

Westland-corp (Lomé, Togo) — info@westland-corp.com