yat-linguoner-ner-model

LinguoNER: Yambeta Named Entity Recognition Model

Model Description

This repository contains LinguoNER, a Transformer-based Named Entity Recognition (NER) model fine-tuned for Yambeta (yat), a low-resource Bantu language spoken in Cameroon.

The model is trained using a silver-standard automatically annotated corpus derived from the Yambeta New Testament and validated through expert-in-the-loop evaluation on sampled subsets.

• Task: Named Entity Recognition (NER)
• Language: Yambeta (yat)
• Entity Types: PER, LOC, ORG
• Model Architecture: Token classification head on top of a Transformer encoder

Base Model

• Base checkpoint: bert-base-cased
• Tokenizer: DS4H-ICTU/yat-bert-tokenizer (WordPiece, Yambeta-specific)

The token-classification head was randomly initialized and fine-tuned jointly with the encoder, following standard practice for NER adaptation.

Training Data

• Dataset: DS4H-ICTU/yat-ner-dataset
• Annotation type: Silver (dictionary-based BIO tagging)
• Gold validation: 500 sentences (annotation logs) + 200 sentences (NER output), reviewed by a domain expert

Dataset Split

• Train: 6317 sentences
• Validation: 790 sentences
• Test: 790 sentences

Evaluation Results (Test Set)

• precision: 0.9885
• recall: 0.9810
• f1: 0.9847
• accuracy: 0.9997

Metrics are reported at token level using standard Precision / Recall / F1.

Intended Use

• Proof-of-concept NER for extremely low-resource African languages
• Baseline for further expert annotation and domain expansion
• Demonstration of Hugging Face workflows under data scarcity

Limitations

• The corpus is Bible-derived, with repetitive narrative structure and limited entity inventory.
• Results should be interpreted as restricted-domain proof-of-concept performance.
• Dictionary-driven annotation may introduce label bias; expert validation mitigates but does not eliminate this.

Citation

If you use this model, please cite:

@misc{linguoner_yambeta_ner,
  title        = {LinguoNER: Yambeta Named Entity Recognition Model},
  author       = {DS4H-ICTU Research Group},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {Model},
  url          = {https://huggingface.co/DS4H-ICTU/yat-linguoner-ner-model}
}