Abstract—With the growth of user-generated content on platforms like social networks and e-commerce, effective sentiment analysis has become increasingly important. This paper presents a comparative study of deep learning models for sentiment analysis, focusing on both Convolutional Neural Networks (CNNs) and prompt-based Bidirectional Encoder Representations from Transformers (BERT) models. We evaluated three CNN variants and multiple prompt strategies for BERT (hand-crafted, adaptive, and hybrid) across five diverse datasets. The results show that the hybrid Gated Recurrent Units_Attention Mechanism (GRU_ATT) BERT model outperforms all others. Prompt-based approaches, particularly adaptive and hybrid designs, offer strong performance and surpass CNN-Non-Static. Moreover, dataset characteristics, such as input length, class distribution, and language structure, significantly impact the performance of the models. Lastly, practical guidelines are proposed for selecting appropriate models based on dataset characteristics.
 
Keywords—Convolutional Neural Network (CNN), Bidirectional Encoder Representations from Transformers (BERT), prompting, sentiment analysis


Copyright © 2025 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).