UzbekTaggerBERT: A Python library for contextual part-of-speech tagging of Uzbek

UzbekTaggerBERT is an open-source Python library for word-level part-of-speech tagging of Uzbek text using a fine-tuned Tahrirchi-BERT model. The package normalizes common apostrophe variants, extracts word spans, aligns Transformer subword predictions through character-span overlap, aggregates overlapping logits, and deterministically labels punctuation. It provides Python and command-line interfaces with slash, JSON, and CoNLL-like outputs. Five-fold cross-validation on UzbekPOS yielded mean accuracy and weighted F1 of 0.9784 ± 0.0023 and 0.9784 ± 0.0023, respectively; the best validation fold (Fold 4) reached 0.9810 and 0.9811. The release includes documentation, tests, benchmarks, PyPI distribution, and a Zenodo archive. • Open-source Python library provides word-level Uzbek POS tagging. • Character-span overlap aligns BERT subwords with original words. • Normalization and punctuation rules improve robust Uzbek inference. • Five-fold mean accuracy and weighted F1 were 0.9784 ± 0.0023. • PyPI, tests, CI, benchmarks and Zenodo support reproducible reuse.

Hali tarjima qilinmagan