ARTIFICIAL INTELLIGENCE-BASED SEMANTIC AND SENTIMENT ANALYSIS OF UZBEK-LANGUAGE ONLINE REVIEWS ON TOURISM SERVICES.

Abstract

The rapid proliferation of online review platforms has created unprecedented volumes of user-generated content in low-resource languages, including Uzbek. This paper presents a comprehensive framework for the semantic and sentiment analysis of Uzbek-language tourist reviews using transformer-based architectures and classical machine learning approaches. We assembled a novel corpus of 14,872 tourism reviews collected from TripAdvisor, Google Maps, and local Uzbek tourism portals covering hotels, restaurants, cultural heritage sites, and guided tours across Tashkent, Samarkand, Bukhara, and Khiva. We fine-tuned a multilingual BERT model (mBERT) and trained a dedicated UzBERT variant on our domain-specific data. In addition, we implemented classical baselines including Support Vector Machines (SVM), Naive Bayes, and LSTM networks. Experimental results demonstrate that UzBERT achieves the highest sentiment classification F1-score of 0.913 on a three-class (positive, neutral, negative) schema, outperforming mBERT by 4.2 percentage points and SVM by 11.7 points. Aspect-level sentiment analysis (ALSA) further identified six prominent tourism service dimensions: accommodation quality, food and beverage, transportation, cultural heritage experience, staff hospitality, and price–value ratio. Key semantic themes were extracted using Latent Dirichlet Allocation (LDA) and validated through qualitative human assessment. Our findings provide actionable insights for Uzbek tourism stakeholders and contribute a reproducible pipeline for NLP in morphologically rich, low-resource Turkic languages.