You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Twi-English Parallel Paragraph Dataset

Dataset Description

This is a parallel Twi-English dataset designed for training machine translation models that understand paragraph structure, including empty lines between text blocks and multiple consecutive sentences.

Dataset Summary

  • Source Language: English
  • Target Language: Twi (Akan)
  • Dataset Structure: Parallel paragraphs with 3 sentences each
  • Paragraph Patterns:
    • 50% with 1 sentence + blank line + 2 sentences
    • 50% with 2 sentences + blank line + 1 sentence

Source Data

The source dataset consists of Ghanaian news articles translated to Twi using the NLLB (No Language Left Behind) 3.3B model. The original English sentences were professionally sourced from Ghanaian news outlets and translated using state-of-the-art neural machine translation.

Data Fields

  • ENGLISH: English paragraph (3 sentences with blank line separator)
  • TWI: Twi paragraph (3 sentences with blank line separator)

Data Splits

This dataset contains a single split with the training data.

Dataset Creation

Curation Rationale

This dataset was created to train MT models that can:

  1. Handle paragraph-level translation
  2. Preserve document structure including blank lines
  3. Understand context across multiple sentences
  4. Maintain proper formatting in translation output

Source Data

  • Initial Data Collection: Ghanaian news articles in English
  • Translation Method: NLLB 3.3B model
  • Processing: Grouped into 3-sentence paragraphs with alternating blank line patterns

Considerations for Using the Data

Discussion of Biases

Since this dataset is machine-translated using NLLB 3.3B, it may contain translation errors or biases present in the original model. Users should be aware that:

  • Machine translation quality may vary
  • Domain-specific terminology may not be perfectly translated
  • Cultural context may be lost in translation

Other Known Limitations

  • Translations are machine-generated and not human-verified
  • Limited to news domain
  • Paragraph structure is artificially created from sentence pairs

Additional Information

Licensing Information

This dataset is released under CC-BY-4.0 license.

Citation Information

If you use this dataset, please cite:

@dataset{twi_english_paragraph_dataset,
  title={Twi-English Parallel Paragraph Dataset},
  author={Mich-Seth Owusu},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/ghananlpcommunity/twi_english_paragraph_dataset}
}
Downloads last month
45