Sentiment Classification of Product Reviews

The rapid growth of e-commerce platforms has led to a significant increase in the volume of user-generated reviews, making manual analysis of customer feedback impractical. A critical issue in modern marketplaces is the presence of inconsistencies between the textual content of reviews and the numerical ratings assigned by users, which can distort product evaluation and mislead potential buyers. This study addresses the task of automatic sentiment classification of textual reviews to identify such discrepancies. A dataset of 100,000 user reviews from the Wildberries marketplace was collected and preprocessed. A three-class sentiment labeling scheme was applied, categorizing reviews as negative, neutral, or positive based on user ratings. Three approaches to sentiment analysis were implemented and compared: logistic regression with TF-IDF vectorization, a transformer-based BERT model, and a hybrid RuBERT model augmented with convolutional neural networks. Experimental results demonstrate that transformer-based and hybrid neural architectures significantly outperform classical machine learning methods, with the RuBERT-CNN model achieving the highest classification performance. The proposed approach can be effectively applied to improve the reliability of rating systems and enhance moderation and analytics tools in e-commerce platforms.