nazarioz/changana-pt-parallel

Portuguese–Changana Parallel Corpus

Description

The first publicly available sentence-level parallel corpus for Portuguese and Changana (Xichangana/Ronga), a Bantu language spoken by approximately 3–5 million people in southern Mozambique, South Africa, and Zimbabwe.

Source

The corpus was constructed by aligning two Bible translations at the verse level:

• Portuguese: Almeida Corrigida Fiel (ACF), published by Sociedade Bíblica Trinitária do Brasil.
• Changana: Bibele hi Xizronga xa Namunhla (BRN), published in 2021 by Dumbeka Editores e Consultores (brn.xizronga.org).

The corpus covers the 27 books of the New Testament.

Statistics

Statistic	Portuguese	Changana
Aligned pairs	7,929	7,929
Mean sentence length (tokens)	20.0	18.4
Vocabulary size	18,715	25,727
CG/PT length ratio	0.92	0.92

Splits

Split	Pairs
Train	6,343
Dev	793
Test	793

Splits created with random shuffle (seed=42).

Languages

• Portuguese (pt): Brazilian Portuguese, formal biblical register.
• Changana (ts): Also known as Xichangana, Xangana, or Ronga. Bantu language (Niger-Congo family). ISO 639-1: ts (Tsonga cluster).

Limitations

• Single domain (religious text)
• Formal/archaic register, not conversational
• Small size (7,929 pairs) by modern MT standards

Usage

from datasets import load_dataset

dataset = load_dataset("nazarioz/changana-pt-parallel")

# Access a training example
example = dataset["train"][0]
print(f"PT: {example['text_pt']}")
print(f"CG: {example['text_cg']}")

Citation

If you use this dataset, please cite:

@misc{changana-pt-parallel-2026,
  title={Portuguese--Changana Parallel Corpus},
  author={Nazario},
  year={2026},
  url={https://huggingface.co/datasets/nazarioz/changana-pt-parallel}
}

License

CC-BY-4.0