EVALUATING CUSTOMER EXPERIENCE IN E-COMMERCE: MULTILINGUAL SENTIMENT ANALYSIS OF USER REVIEWS USING TRANSFORMER MODELS

Abstract

This study addresses the challenge of multilingual sentiment analysis in e-commerce, with a focus on Ukrainian and Russian book reviews. We propose a hybrid framework based on transformer architectures that accounts for the linguistic complexity of Slavic languages and the significant class imbalance often present in customer feedback data. Using a dataset of approximately 70,000 user reviews from the Ukrainian online bookstore Yakaboo, we evaluate four model variants based on the XLM-RoBERTa architecture and compare their performance to a monolingual Ukrainian RoBERTa baseline. The classification task is formulated as binary: identifying low-rated reviews (1–3 stars) versus high-rated ones (4–5 stars), with the minority class comprising only 4–5% of the dataset. Our best-performing model — combining partial fine-tuning, a deep classifier, and focal loss — achieves a macro F1 score of 0.73 and an F1 score of 0.47 for the minority class, outperforming the baseline (0.66 and 0.37, respectively). The application of oversampling and focal loss proved effective in mitigating class imbalance. Beyond technical performance, the findings underline the practical utility of accurate sentiment detection for e-commerce platforms. Effective identification of negative feedback enables companies to address customer dis- satisfaction, refine product offerings, and inform personalized marketing strategies. This research contributes to advancing multilingual NLP in low-resource settings and provides a scalable solution for real-world sentiment classification in morphologically rich languages. The scientific novelty of this research lies in the integration of multilingual transformer architectures with adaptive classifiers and class imbalance mitigation techniques, specifically tailored for real-world e-commerce review data in morphologically rich languages. The proposed framework demonstrates how domain-specific fine-tuning and architectural customization can significantly enhance sentiment classification performance in low-resource language settings.