TY - JOUR

T1 - UzNER: A Human-Reviewed Benchmark for Uzbek Named Entity Recognition With Gazetteer-Augmented Transformer Models

AU - Saidov, Bobur

AU - Barakhnin, Vladimir

AU - Fayzullaeva, Zarnigor

AU - Ibragimov, Umid

AU - Tursunov, Ulugbek

N1 - Saidov, B., Barakhnin, V., Fayzullaeva, Z., Ibragimov, U. & Tursunov, U. (2026). UzNER: A Human-Reviewed Benchmark for Uzbek Named Entity Recognition With Gazetteer-Augmented Transformer Models. Journal of Computer Science, 22(6), 1894-1911. https://doi.org/10.3844/jcssp.2026.1894.1911 This research was partially supported by the state assignment of the Ministry of Science and Higher Education of the Russian Federation for the Federal Research Center for Information and Computational Technologies. The funding number is not applicable/was not provided for this state assignment. The APC was selffunded by the authors.

PY - 2026/6/22

Y1 - 2026/6/22

N2 - UzNER-100K is a large-scale human-reviewed benchmark for Uzbek named entity recognition with 100,000 training sentences, 18 fine-grained entity types and 200,083 entity mentions across 114,269 sentences in total. The corpus was constructed through an LLM-assisted, expert-reviewed annotation pipeline that achieved strong reliability on the main audit subset while substantially reducing corpus-construction effort. The benchmark includes a standard test split, a gold-audited subset and a hard subset designed to stress long, ambiguous and structurally complex cases. We evaluate 10 Uzbek NER systems spanning recurrent, monolingual Uzbek, multilingual transformer and hybrid architectures. The best model, XLM-R + Gazetteer + CRF, reaches 91.03 Micro-F1 on the standard test set, 89.67 on the gold-audited subset and 83.21 on the hard subset. Quality control included a dedicated inter-annotator agreement audit, achieving 91.3% span-level agreement, 93.7% entity-type agreement, and a Cohen’s Kappa of 0.914. In addition, a qualitative native-speaker assessment confirmed the linguistic naturalness of the model outputs while highlighting remaining challenges in legal, administrative, and event-related expressions.

AB - UzNER-100K is a large-scale human-reviewed benchmark for Uzbek named entity recognition with 100,000 training sentences, 18 fine-grained entity types and 200,083 entity mentions across 114,269 sentences in total. The corpus was constructed through an LLM-assisted, expert-reviewed annotation pipeline that achieved strong reliability on the main audit subset while substantially reducing corpus-construction effort. The benchmark includes a standard test split, a gold-audited subset and a hard subset designed to stress long, ambiguous and structurally complex cases. We evaluate 10 Uzbek NER systems spanning recurrent, monolingual Uzbek, multilingual transformer and hybrid architectures. The best model, XLM-R + Gazetteer + CRF, reaches 91.03 Micro-F1 on the standard test set, 89.67 on the gold-audited subset and 83.21 on the hard subset. Quality control included a dedicated inter-annotator agreement audit, achieving 91.3% span-level agreement, 93.7% entity-type agreement, and a Cohen’s Kappa of 0.914. In addition, a qualitative native-speaker assessment confirmed the linguistic naturalness of the model outputs while highlighting remaining challenges in legal, administrative, and event-related expressions.

KW - Benchmark Dataset

KW - Gazetteer-Enhanced Decoding

KW - Low-Resource NLP

KW - Multilingual Transformers

KW - Uzbek NER

UR - https://www.scopus.com/pages/publications/105044228554

UR - https://www.mendeley.com/catalogue/19513578-c6be-3003-b4d9-01cd02e32276/

U2 - 10.3844/jcssp.2026.1894.1911

DO - 10.3844/jcssp.2026.1894.1911

M3 - Article

VL - 22

SP - 1894

EP - 1911

JO - Journal of Computer Science

JF - Journal of Computer Science

SN - 1552-6607

IS - 6

ER -