GinLishCorpus-bi-v1.0

Dataset Summary

GinLishCorpus-bi-v1.0 is a high-quality bidirectional English–Tagin parallel corpus developed for training and evaluating bi-directional Neural Machine Translation (NMT) systems.
The corpus is specifically designed to support low-resource language research, with a focus on the Tagin language (Tani language family).

The dataset supports both translation directions:

• English → Tagin
• Tagin → English

This corpus is suitable for fine-tuning and training multilingual and bilingual NMT models such as mBART50, MarianMT, M2M-100, and Transformer-based architectures.

---

Languages

• English (en)
• Tagin (tgj) – written in a modified Roman script

---

Dataset Structure

The dataset is provided in parallel sentence format, where each entry contains aligned sentence pairs.

Format

Typical file formats include:

• JSON

Columns

Column Name	Description
tagin	Tagin sentence (source or target)
english	English sentence (source or target)

The same parallel data can be used in both directions during training by swapping source and target languages.

---

Domain Coverage

GinLishCorpus-bi-v1.0 covers multiple domains to enhance translation robustness:

• Daily conversation
• Social and informal communication
• Religious and biblical texts
• Culturally grounded expressions
• Common narrative and instructional sentences

---

Data Collection and Creation

• Sentences were manually curated, translated, and verified
• Additional parallel data was generated using transfer learning and NMT-assisted augmentation
• Quality control involved manual validation and automatic filtering
• Emphasis was placed on semantic faithfulness and cultural adequacy

---

Intended Uses

Primary Use

• Training bi-directional English ↔ Tagin NMT models
• Fine-tuning pre-trained multilingual models
• Benchmarking low-resource MT systems

Secondary Use

• Cross-lingual transfer learning
• Data augmentation for Tagin NLP tasks
• Curriculum learning for multilingual MT
• Linguistic and typological analysis

---

Out-of-Scope Uses

• High-stakes decision-making (e.g., legal, medical)
• Production systems without additional validation
• Fully representative sociolinguistic modeling of Tagin

---

Dataset Size

• Approximately 150K parallel sentence pairs
• Balanced across translation directions

---

Evaluation

The dataset has been successfully used to train and evaluate:

• repleeka/ginlishMT-mbart50-tgj-en-bi fine-tuned model
• A Transformer-based bilingual NMT system

Automatic evaluation metrics such as BLEU, chrF, and TER indicate strong performance for a low-resource language pair.

---

Ethical Considerations

• No personally identifiable information (PII) is included
• Texts are derived from public, neutral, or researcher-generated sources
• Cultural and religious content is handled with care and respect

---

Bias and Limitations

• Domain bias toward conversational and religious text
• Limited coverage of highly technical or specialized domains
• Romanized Tagin script may differ from alternative orthographic conventions

---

Citation

If you use this dataset, please cite it as:

@dataset{ginlishcorpus_bi_v1,
  title     = {GinLishCorpus-bi-v1.0: A Bidirectional English--Tagin Parallel Corpus},
  author    = {Tungon Dugi},
  year      = {2026},
  version   = {1.0},
  publisher = {Hugging Face},
}

Contact

For questions, issues, or collaboration requests, please contact the dataset maintainer via the associated Hugging Face repository.

Acknowledgements

We acknowledge the contributions of:

• Native speakers and annotators
• Open-source NLP and MT communities
• Multilingual pre-trained model developers

This dataset aims to advance Tagin language technology and promote inclusive NLP research for low-resource languages.