Multi-Class Sentiment Analysis on Turkish E-Commerce Data: Comparison of Traditional and Deep Learning Approaches

In this study, traditional machine learning and deep learning approaches are compared for multi-class sentiment analysis on Turkish e-commerce reviews. A balanced dataset (36,359 positive, 36,359 neutral, 36,359 negative) was collected from T & uuml;rkiye's leading e-commerce platforms via web scraping and labeled with Label Studio. Experimental results show that deep learning models outperform traditional methods. LSTM-based models effectively captured long-term dependencies, while hybrid CNN+LSTM (95.0%) and CNN+GRU (94.9%) architectures, as well as bidirectional models BiLSTM (94.2%) and BiGRU (94.2%), achieved strong classification performance. The highest accuracy was obtained with the LSTM model (95.6%). When the model performances were examined, the neutral class was predicted with higher accuracy than the positive and negative classes. This may be explained by the more distinctive and less ambiguous linguistic patterns of neutral reviews, whereas positive and negative expressions often overlap semantically and are highly context-dependent. Overall, this research contributes to the determination of the most suitable deep learning models for multi-class sentiment analysis in the field of Turkish natural language processing, while revealing that hybrid and bidirectional architectures in particular offer an effective alternative. The findings also pave the way for new research aimed at eliminating the uncertainties experienced in transitions between classes.