Skip to content
DSML.KZ
Publications

Talk · Ural Symposium on Biomedical Engineering, Radioelectronics and Information Technology · May 14, 2026

Entity-Aware Sentiment Analysis for Uzbek Social Media Using XLM-RoBERTa NER and LSTM

Bobur Rashidovich Saidov, Zarnigor Fayzullaeva, Ulugbek Tursunov, Zayniddin Narmuratov, Shokhida Jalilovna Egamova, Begmuratova A. Gulbakhar

Bobur Saidov
Open via DOI

Abstract

This paper studies how named entities influence sentiment polarity in Uzbek social media and proposes an entity-aware sentiment analysis pipeline. We collected $\mathbf{1 0, 0 0 0}$ Uzbek posts from Telegram, Instagram, and Facebook (January-June 2025) and annotated three entity types (Person, Brand, Geolocation) in a subset of $\mathbf{4, 6 0 0}$ posts. Named entities are extracted with an adapted XLM-RoBERTa NER model, while sentiment is predicted by an LSTM classifier with Uzbek Word2Vec embeddings. To account for the different impact of entity types, we introduce an entity-weighted adjustment that calibrates sentiment scores using empirically estimated coefficients. On the held-out test split, the proposed approach achieves strong sentiment performance, reaching Test Accuracy=0.84 and Macro-F1=0.83 (Macro Precision=0.83, Macro Recall = 0.82). Statistical validation confirms significant differences across entity categories (ANOVA) and a positive association between entity mentions and model confidence (Pearson $\mathbf{r}=$ 0.82). The approach supports practical analytics scenarios such as brand reputation monitoring, public communication tracking, and tourism feedback analysis for Uzbek-language data.

Sentiment Analysis and Opinion MiningTopic ModelingHate Speech and Cyberbullying Detectionsentiment analysisnamed entity recognitionUzbek language

More by Bobur Saidov