TatarNLPWorld/tatar-morphology-benchmark

Tatar Morphology Benchmark

This repository contains evaluation results for morphological analysis models trained on the Tatar Morphological Corpus.

Models Evaluated

• mBERT
• RuBERT
• DistilBERT
• LSTM
• Turkish BERT
• XLM-R

Key Results (Test Set Accuracy)

Model	Accuracy	F1 (micro)
mBERT	0.9905	0.9905
RuBERT	0.9861	0.9861
DistilBERT	0.9850	0.9850
XLM-R	0.9837	0.9837
LSTM	0.9440	0.9440
Turkish BERT	0.8769	0.8769

All results are based on a test set of 7999 sentences (80k training). Detailed metrics with 95% confidence intervals are available in final_results_with_ci.csv.

Contents

• final_results_with_ci.csv – main metrics with confidence intervals
• pos_accuracy.csv – accuracy per part-of-speech
• model_comparison.png – bar chart comparison
• training_curves_combined.png – loss/accuracy curves
• confusion_matrices.png – confusion matrices for top models
• accuracy_by_length.csv – accuracy grouped by sentence length
• tag_frequencies.csv – tag frequency statistics

Citation

If you use these results, please cite:

@misc{tatar-morph-benchmark,
  author = {Arabov Mullosharaf Kurbonovich, TatarNLPWorld},
  title = {Tatar Morphology Benchmark},
  year = {2026},
  howpublished = {Hugging Face Dataset},
  url = {https://huggingface.co/datasets/TatarNLPWorld/tatar-morphology-benchmark}
}