Hybrid BERT–SVM Model for Three-Class Sentiment Analysis of Multi-Domain Uzbek Online Texts

Sentiment analysis remains challenging for low-resource and morphologically rich languages because classification performance depends on both representative annotated data and effective contextual text representations. This study develops a hybrid BERT–SVM architecture for three-class sentiment analysis of Uzbek online texts. The proposed model combines task-adapted contextual representations from bert-base-multilingual-cased with the maximum-margin classification capability of a Linear Support Vector Machine. Multilingual BERT was fine-tuned on the target sentiment corpus, after which the final-layer [CLS] representation was extracted as a 768-dimensional contextual feature vector and supplied to a One-vs-Rest Linear SVM classifier. The experimental corpus contained 7,963 manually annotated Uzbek texts, including 2,987 Positive, 2,419 Negative, and 2,557 Neutral observations. The data were collected from social media platforms, news websites, Google Play, and Google Maps and covered restaurant, medicine, hotel and tourism, and education domains. Two annotators participated in labeling, and Cohen's kappa coefficient was 0.83 ( = 0.83). The model was evaluated using stratified five-fold cross-validation with aggregated out-of-fold predictions so that each observation was evaluated by a model that had not used that observation during the corresponding training iteration. The hybrid architecture correctly classified 6,879 of 7,963 observations, producing an overall Accuracy of 86.39%, Macro-Precision of 86.23%, Macro-Recall of 86.29%, and Macro-F1 score of 86.26%. Class-level F1-scores were 88.37% for Positive, 84.84% for Negative, and 85.57% for Neutral sentiment. Confusion-matrix analysis showed that 68.63% of all classification errors involved the Neutral category. The findings demonstrate that fine-tuned multilingual BERT representations can provide a discriminative contextual feature space for subsequent maximum-margin classification and establish a hybrid baseline for further comparative and domain-aware Uzbek sentiment analysis. A TF-IDF + Linear SVM baseline achieved 73.00% ± 1.09% Accuracy and 72.34% ± 1.14% Macro-F1 across five folds. The corpus was assembled during 2023–2024 through a combination of automated, semi-automated, and manual collection procedures.