PrinceAlhassanNasamu/kusaal-english-parallel-corpus

Kusaal-English Parallel Corpus

The first open parallel corpus for Kusaal — a Gur language spoken by ~400,000 people in northern Ghana and parts of Burkina Faso. Kusaal has no entry in Google Translate, no presence in Meta's NLLB-200, and no prior open NLP dataset.

This corpus was assembled from scratch by a native Kusaal speaker from Bawku, Ghana, and used to train the first open-source Kusaal-English machine translation model: PrinceAlhassanNasamu/kusaal-nllb-600M.

---

Dataset Details

Dataset Description

34,568 Kusaal-English sentence pairs drawn from five sources and expanded through back-translation augmentation. Covers religious text, daily life, health, agriculture, vocabulary/dictionary entries, and general reference.

• Language pair: Kusaal (kus_Latn) ↔ English (eng_Latn)
• Script: Latin
• Total pairs: 34,568
• License: CC BY 4.0
• Curated by: Prince Nasamu Alhassan

---

Sources

Source	Pairs	Domain
YouVersion (Bible KJV)	~29,257	Religious / formal
English-Kusaal Index	3,504	Index / vocabulary
GhanaNLP	3,489	Daily life, health, agriculture, greetings, numbers
Lexique Pro (Kusaal lexical database)	2,775	Dictionary
Wikipedia	1,136	General
Back-translation augmentation	~2,407	Mixed
Total	~34,568

---

Data Splits

Split	Pairs
Train	~27,300
Validation	~5,187
Test	~2,081

---

Dataset Structure

Data Fields

• kusaal — sentence in Kusaal (Latin script)
• english — corresponding English translation

Data Instances

{"kusaal": "Laafi bɛ?", "english": "Are you all right?"}
{"kusaal": "Mam yɛla bʋ ŋa.", "english": "I do not understand."}

---

Data Collection and Processing

Collection

Data was sourced from publicly available texts and community linguistic resources:

• YouVersion Bible (KJV): Scraped and aligned at the verse level with the Kusaal Bible text
• GhanaNLP: Parallel sentences from GhanaNLP's published Kusaal dataset
• English-Kusaal Index: Structured bilingual index entries
• Lexique Pro: Exported from the locally installed Kusaal lexical database
• Wikipedia: Machine-translated and manually reviewed general-domain sentences
• Back-translation: Monolingual Kusaal text translated to English using an intermediate model checkpoint, filtered by confidence and length ratio

Cleaning Pipeline

Two significant data quality issues were identified and fixed before training:

1. HTML pollution — YouVersion scraper output included raw HTML markup (class="verse v2" > 2 In those days...) in thousands of pairs. Fixed by stripping HTML patterns and re-sourcing clean English text from the KJV JSON.

2. CSV quoting corruption — Unescaped apostrophes in Bible text caused pandas to silently drop thousands of rows. Fixed by re-saving all source files with QUOTE_ALL.

Post-cleaning steps: deduplication on Kusaal+English pairs, length ratio filter (max 6×), UTF-8 encoding throughout.

---

Limitations

• Domain bias: ~85% of pairs originate from Bible text. The corpus is stronger in formal and religious Kusaal than in casual speech.
• Synthetic pairs: Back-translated pairs introduce translation errors from the intermediate model. These are included at a controlled ratio.
• Tokenization: Kusaal is not in NLLB's SentencePiece vocabulary. Models trained on this corpus will over-segment Kusaal words into subword fragments.
• No audio: This is a text-only parallel corpus. An audio-aligned version is planned.
• No native speaker annotation: Pairs have not been formally validated by trained Kusaal linguists, though the curator is a native Kusaal speaker from Bawku.

---

Uses

Direct Use

• Training or fine-tuning machine translation models for Kusaal-English
• Evaluating existing multilingual models on Kusaal
• Linguistic analysis of Kusaal sentence structure

Out-of-Scope Use

• Pairs from Bible text are formal and should not be treated as representative of conversational Kusaal

---

Citation

@misc{alhassan2026kusaal,
  author    = {Alhassan, Prince Nasamu},
  title     = {Kusaal-English Parallel Corpus},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/PrinceAlhassanNasamu/kusaal-english-parallel-corpus}
}

---

About

Built by Prince Nasamu Alhassan — native Kusaal speaker from Bawku, Ghana. Mathematical Sciences with Computer Science student at the University of Ghana, Legon. Independent NLP researcher affiliated with GhanaNLP.

• Model: PrinceAlhassanNasamu/kusaal-nllb-600M
• GitHub: @NasamuAlhassan
• Email: pnalhassan@gmail.com