Sentiment Analysis for Uzbek Language Using Pre-trained Multilingual Models
Year: 2026
Type: conference-paper
Abstract: Uzbek sentiment analysis has been held back for years, mostly because there were no public datasets, no shared benchmarks, and no tools built specifically for the language. We wanted to fix that. We collected and annotated a multi-domain corpus of 10,000 sentences labeled for polarity, plus a 3,000-sentence subset for aspect-based sentiment, drawing on posts from Twitter/X, Telegram, and Uzbek news comment sections. This is the first fully reproducible benchmark for the language. On top of this data, we fine-tuned four multilingual transformer models (mBERT, XLM-R base, XLM-R large, and mT5 base) under the same training conditions, and we carefully measured how much each preprocessing step helps: script normalization, morphological pre-segmentation, back-translation, and lexical substitution. XLM-R large gave us the best numbers, reaching 84.7% accuracy and 82.9% macro-F1 on the sentence-level task, which is up to 22.7 percentage points above anything previously reported for Uzbek. The biggest single gain came from normalizing the mixed Cyrillic-Latin script, which alone added 5.4 percentage points. Combining augmentation techniques pushed the total improvement to 5.8 points. The errors we still see cluster around a few stubborn problems: neutral-positive confusion, code-switching inside a sentence, negation that stretches across many words, and sarcasm. The practical message is straightforward: when you're working with a low-resource language like Uzbek, language-aware preprocessing is not something you can skip; it is absolutely central to getting decent results.
Authors: Dilnoza Boltayeva, Gulnoza Salimova
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224891576