Sentiment analysis of Uzbek texts using NER: a comparative study of SVM, LSTM, and BERT models

Abstract

This paper presents a comparative analysis of machine learning (SVM), deep learning (LSTM), and transformer-based (BERT) models for sentiment classification in Uzbek texts, enhanced by Named Entity Recognition (NER). The study addresses the challenge of accurately detecting sentiment in morphologically complex languages with limited resources, focusing on Uzbek–a Turkic language with rich agglutinative structures. A dataset of 10,000 user-generated comments from social platforms was annotated using a hybrid approach: manual labeling for sentiment (positive, negative, neutral) and a CRF-based NER system to identify entities (e.g., brands, locations, public figures). The integration of NER features aimed to resolve contextual ambiguities, such as distinguishing between "I love Samarkand’s history" (positive) and "Samarkand’s traffic is unbearable" (negative). Experimental results demonstrate that BERT, fine-tuned on Uzbek text, achieved the highest accuracy (90.2%) by leveraging contextualized embeddings to align entities with sentiment. LSTM showed competitive performance (85.1%) in sequential pattern learning but required extensive training data. SVM, while computationally efficient, lagged at 78.3% accuracy due to its inability to capture nuanced linguistic dependencies. The findings emphasize the critical role of NER in low-resource languages for disambiguating sentiment triggers and propose practical guidelines for deploying BERT in real-world applications, such as customer feedback analysis. Limitations, including data scarcity and computational costs, are discussed to inform future research on optimizing lightweight models for Uzbek NLP tasks.

More by Bobur Saidov

• Human-Reviewed Uzbek Legal Named Entity Recognition DatasetF1000Research · 2026
• UzNER: A Human-Reviewed Benchmark for Uzbek Named Entity Recognition With Gazetteer-Augmented Transformer ModelsJournal of Computer Science · 2026 · Citations: 3
• Entity-Aware Sentiment Analysis for Uzbek Social Media Using XLM-RoBERTa NER and LSTMUral Symposium on Biomedical Engineering, Radioelectronics and Information Technology · 2026
• Software System for Hybrid Sentiment Analysis of Uzbek Texts using Named EntitiesThe Herald of the Siberian State University of Telecommunications and Information Science · 2026