<![CDATA[The rapid growth of electronic commerce platforms has generated a large volume of user-generated product reviews, increasing the need for accurate sentiment analysis. This study proposes a hybrid deep learning architecture that combines a sentiment-optimized transformer, SiEBERT, as a frozen encoder with a Long Short-Term Memory (LSTM) network as a sequence classifier. The model is evaluated on an Amazon review dataset containing 34,626 instances labeled into positive and negative sentiments. Three configurations are compared: SiEBERT as a direct classifier, LSTM trained from scratch, and the proposed hybrid model. Experimental results show that SiEBERT achieves 73.72% accuracy, while LSTM reaches 57.00%. The hybrid model achieves the best performance with 93.76% accuracy and a weighted F1 score of 94.15%. These findings indicate that combining contextual embeddings with sequential modeling produces more effective sentiment representations. This study contributes by introducing a hybrid SiEBERT–LSTM architecture that addresses limitations of standalone models and significantly improves classification performance, particularly in handling imbalanced sentiment data]]>
Copyrights © 2026
