uzbek-tagger-bert 0.1.1

Part-of-Speech Tagger for the Uzbek Language based on Tahrirchi-BERT

====================================================================== UzbekTaggerBERT: High-Accuracy POS Tagger for the Uzbek Language

UzbekTaggerBERT is an open-source Part-of-Speech (POS) tagging library designed specifically for the highly agglutinative Uzbek language. It is built by fine-tuning the state-of-the-art Tahrirchi-BERT (RoBERTa architecture) on a comprehensive Uzbek morphological dataset.

This project solves major tokenization challenges inherent in agglutinative languages, providing researchers and developers with a robust, highly accurate tool for Natural Language Processing (NLP) in Uzbek.

MODEL PERFORMANCE & HUGGING FACE

The core model used in this tagger is publicly available on the Hugging Face Hub: URL: https://huggingface.co/MaksudSharipov/Uzbek-POS-Tagger-TahrirchiBERT

Evaluation metrics on the UzbekPOS test dataset demonstrate state-of-the-art results:

• Accuracy: 0.9810
• Weighted F1: 0.9811

KEY SCIENTIFIC FEATURES

• Contextual Homonym Disambiguation: Accurately distinguishes between morphologically identical but semantically different words based on context (e.g., "yuz" as a NOUN vs. "yuz" as a NUM).
• Smart Text Normalization: Automatically standardizes various Cyrillic/Latin apostrophes to prevent the tokenizer from incorrectly fragmenting words like "O'zbekiston".
• Overlap-based Offset Mapping: Employs an advanced mathematical overlap algorithm to perfectly align model-generated subword tokens back to their original root words, eliminating the "UNKNOWN" tag issue common in Transformer-based pipelines.

INSTALLATION

You can install the package directly from PyPI using pip:

pip install uzbek-tagger-bert

Dependencies: This package requires "transformers" and "torch". They will be installed automatically if not already present.

QUICK START & EXAMPLE

Using the tagger is straightforward. It abstracts away the complexities of subword tokenization and tensor operations.

from uzbek_tagger_bert import UzbekTaggerBERT

# 1. Initialize the tagger
tagger = UzbekTaggerBERT()

# 2. Input a complex sentence in Uzbek
text = "U yuz burib ketdi va yuz soat ishladi.\nAlisher sariq olma olib menga qizil olma olmading dedi. \nO'zbekiston eng baxtli davlatlar reytingida 53-o'rinda qoldi."

# 3. Get the tagged output
result = tagger(text)

print(result)

[Expected Output]

U/PRON yuz/NOUN burib/VERB ketdi/VERB va/CCONJ yuz/NUM soat/NOUN ishladi/VERB ./PUNCT 
Alisher/PROPN sariq/ADJ olma/NOUN olib/VERB menga/PRON qizil/ADJ olma/NOUN olmading/VERB dedi/VERB ./PUNCT 
O'zbekiston/PROPN eng/ADV baxtli/ADJ davlatlar/NOUN reytingida/NOUN 53-o'rinda/NUM qoldi/VERB ./PUNCT

METHODOLOGY & TOKENIZATION HANDLING

Because Uzbek heavily relies on suffixes, standard tokenizers split words into multiple sub-tokens. This tool overrides standard pipeline constraints and uses dynamic boundary checking to reconstruct original morphological entities seamlessly before assigning the final POS tag.

ACADEMIC CITATION

This tagger model was fine-tuned by expanding the multi-domain dataset introduced in our published research. If you use this library, model, or dataset in your academic and research projects, please strictly cite the following paper:

[APA Style] Sharipov, M., Kuriyozov, E., & Vičič, J. (2026). UzbekPOS: A multi-domain dataset for Uzbek part-of-speech tagging. Data in Brief, 112640. https://doi.org/10.1016/j.dib.2026.112640

[BibTeX]

@article{sharipov2026uzbekpos,
  title={UzbekPOS: A multi-domain dataset for Uzbek part-of-speech tagging},
  author={Sharipov, Maksud and Kuriyozov, Elmurod and Vičič, Jernej},
  journal={Data in Brief},
  pages={112640},
  year={2026},
  publisher={Elsevier},
  doi={10.1016/j.dib.2026.112640}
}

LICENSE

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.