KyrgyzNER: Human-Annotated NER Dataset for Kyrgyz

Overview

KyrgyzNER is the first manually annotated Named-Entity Recognition (NER) dataset for the Kyrgyz language.
It consists of 1,499 news articles (10,900 sentences, 140k tokens) from the 24.kg news portal, annotated with 39,075 entity mentions across 27 classes.

This project provides both the dataset and baseline models, aiming to advance NLP research for low-resource Turkic languages.

---

Key Features

• 📑 Dataset:

  • 1,499 Kyrgyz news articles (2017–2022)
  • 10,900 sentences, 39,075 entity mentions
  • 27 entity categories (Person, Location, Institution, Period, etc.)
  • Format: CoNLL-2003

• 🛠 Annotation:

  • Annotated by 59 trained Kyrgyz linguists and students
  • Guidelines adapted from GROBID-NER
  • High-quality dataset with κ = 0.89 inter-annotator agreement

• 🤖 Models:

  • Baselines: CRF, BiLSTM+CRF, multilingual BERT, mT5
  • Best results: XLM-RoBERTa-base (F1 ≈ 0.73)
  • HuggingFace Model ready to use

---

Usage

You can directly load and use our fine-tuned model from HuggingFace:

from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline

model_name = "TTimur/xlm-roberta-base-kyrgyzNER"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)

ner_pipeline = pipeline("ner", model=model, tokenizer=tokenizer)

text = "Президент Садыр Жапаров бүгүн Бишкекте сүйлөө жасады."
print(ner_pipeline(text))

---

Dataset Statistics

Split	Docs	Sentences	Tokens	Mentions
Train (999)	999	7,033	89,248	24,949
Test (500)	500	3,867	51,118	14,126
Total	1499	10,900	140,366	39,075

• Most frequent classes: Person, Location, Institution, Measure
• Rare classes (few samples): Award, Animal, Substance, Identifier

---

Baseline Results

Model	Precision	Recall	F1
CRF	0.70	0.55	0.62
BERT+CRF	0.67	0.63	0.65
mBERT (cased)	0.68	0.68	0.68
mT5-small	0.70	0.68	0.69
XLM-RoBERTa	0.74	0.71	0.73

---

Contribution

We are grateful to:

• 59 volunteers (mainly students of KSTU) who annotated the dataset
• Dr. Gulnara Kabaeva and Dr. Gulira Zhumalieva for academic support
• List of contributors

We welcome contributions from the NLP community. Please open issues or pull requests if you’d like to improve the dataset, guidelines, or models.

---

Citation

If you use this dataset or model in your research, please cite:

@inproceedings{turatali2025kyrgyzner,
  title     = {Human-Annotated NER Dataset for the Kyrgyz Language},
  author    = {Turatali, Timur and Alekseev, Anton and Jumalieva, Gulira and Kabaeva, Gulnara and Nikolenko, Sergey},
  booktitle = {Proceedings of TurkLang 2025},
  year      = {2025}
}

---

License

• Dataset: CC BY-NC-SA 4.0
• Code & Models: MIT license

---

👉 Full details are available in our paper.