Green Sentiment Analysis in E-Commerce Reviews: A Comparative Machine Learning Approach on Turkish Consumer Feedback

This study investigates the effectiveness of three sentiment classification approaches, i.e. Logistic Regression (LR), Support Vector Classification (SVC), and a fine-tuned BERTurk model, on Turkish e-commerce reviews related to environmentally conscious, or “green,” products. Using a real-world dataset drawn from Trendyol, one of Türkiye’s largest online marketplaces, we preprocessed and filtered the data to focus on user-generated product comments that reference sustainability-oriented themes. Each model was evaluated using standard classification metrics, including accuracy and macro-averaged F1-score, to assess both overall performance and sensitivity to class imbalance. The results show that while classical machine learning methods such as LR and SVC provide reasonably high accuracy, they struggle to distinguish neutral sentiment effectively, which is an issue commonly encountered in Turkish-language sentiment tasks. In contrast, the BERTurk model achieved the highest overall performance, with an accuracy of 0.91 and a macro F1-score of 0.67. It was particularly effective in detecting positive and negative sentiment, while still exhibiting the known difficulty of identifying neutral expressions. These findings suggest that transformer-based models offer a clear advantage in extracting sentiment from morphologically rich languages like Turkish, especially in domains where emotional nuance and linguistic ambiguity are prevalent. The study contributes to both the sentiment analysis literature and Management Information Systems research by demonstrating the value of domain-specific deep learning for consumer analytics in green commerce. It highlights practical implications for businesses aiming to understand and respond to public attitudes toward sustainable products and emphasizes the need for improved modeling of neutral sentiment. Future work should focus on expanding Turkish sentiment datasets, addressing class imbalance, and refining model architectures to better capture the subtleties of eco-conscious consumer expression.