Multi-Class Sentiment Analysis with E-Commerce User Reviews: Comparisons of Classical and Deep Learning Applications

Abstract

The spread of the internet, the increase in online shopping sites, and users sharing their experiences of these products on these sites have led to a significant increase in the amount of text-based data on e-commerce sites. Interpreting this data has made sentiment classification studies important in terms of analyzing user satisfaction. In this study, sentiment classification was performed on 6,367 Turkish product reviews from the clothing category on e-commerce websites. Various filtering steps were applied during the data preprocessing process. These steps directly affected the overall success of the model. During this process, lowercase conversion, punctuation, and special character cleaning steps were performed. Subsequently, traditional machine learning algorithms such as Support Vector Machines (SVM) and Naive Bayes were trained using the BERTurk transformerbased model, and their performances were compared using accuracy and F1 score metrics. In the evaluations, the BERTurk cased 128k model yielded the best results, with an F1 score and accuracy rate of 83.3%, outperforming other models. Based on the results obtained, the BERTurk model showed the highest success, with an accuracy rate of 93.0% in the 'positive' class and 87.5% in the 'negative' class. The 60.7% accuracy rate in the 'neutral' class, however, showed relatively lower performance despite the semantic ambiguities in the dataset and the previous manual label cleaning. The results indicate that transformer-based approaches for Turkish sentiment classification yield better results compared to traditional machine learning methods. © 2025 IEEE.