UzNER-Style(v2)
Description
This dataset presents a multi-genre named entity recognition (NER) corpus for the Uzbek language. The corpus contains 75,000 manually annotated sentences collected from five functional styles of Uzbek: official, journalistic, scientific, literary, and conversational. Each style includes 15,000 sentences, making the dataset suitable for both in-domain NER evaluation and cross-style robustness analysis. The dataset is annotated using the BIOES tagging scheme and covers a broad set of named entity categories, including persons, organizations, geopolitical entities, dates, money expressions, quantities, durations, percentages, locations, laws, documents, positions, events, products, languages, and other domain-relevant entity types. The inclusion of different Uzbek styles allows researchers to evaluate how NER models perform across diverse lexical, syntactic, and contextual conditions. The corpus is intended to support research on Uzbek natural language processing, low-resource NER, sequence labeling, domain adaptation, and cross-genre model generalization. It can be used for training, validation, testing, and comparative evaluation of machine learning and deep learning models for Uzbek named entity recognition.
Files
Steps to reproduce
The dataset was constructed through the following steps. First, Uzbek texts were collected from five functional styles: official, journalistic, scientific, literary, and conversational Uzbek. Each style was represented by 15,000 sentences. The collected texts were cleaned, normalized, and segmented into sentences and tokens. Named entities were then annotated according to the BIOES tagging scheme using a predefined set of entity categories. The annotation process included manual checking and correction to ensure consistency of entity boundaries and labels across styles. The final corpus contains 75,000 annotated sentences and can be used for training, validation, testing, and cross-style evaluation of Uzbek named entity recognition models.