Research Information System
7th International Conference on Intelligent and Fuzzy Systems, INFUS 2025, İstanbul, Turkey, 29 - 31 July 2025, vol.1529 LNNS, pp.37-44, (Full Text)
This research explores deep learning techniques for sentiment classification of product reviews. The dataset, generated by ChatGPT, consists of 5,000 synthetic product reviews labeled as positive (40%), neutral (30%), and negative (30%). A preprocessing pipeline was implemented to standardize sentiment labels and clean textual data for consistency. Four text representation techniques—TF-IDF, Bag of Words (BoW), Word2Vec, and SBERT were used to convert text into numerical features. Additionally, four neural network architectures were applied: a bidirectional LSTM, a CNN-based model, a deep dense network, and a hybrid CNN+LSTM approach. The models were trained using an 80-20 train-test split, with 50 epochs and a batch size of 32. The Adam optimizer facilitated training convergence, and model performance was assessed using accuracy, F1-score, and confusion matrices. The bidirectional LSTM model achieved the highest accuracy at 89.2%, followed by the hybrid CNN+LSTM model at 87.5%. TF-IDF and BoW yielded competitive results, while Word2Vec and SBERT showed slightly lower performance despite their semantic depth, with the deep dense network reaching 83.4% accuracy. These results highlight the importance of feature representation and model selection in sentiment analysis. This study advances natural language processing methodologies for customer feedback analysis, offering insights into the trade-offs between computational efficiency and classification performance.