In recent years, with the rapid impact of the internet economy on the traditional economy, online shopping has become increasingly popular among contemporary consumers compared to offline shopping. Particularly during the pandemic in recent years, when people were restricted from going outside, online shopping became the primary method of purchasing goods. A large volume of e-commerce review data contains consumers’ emotional tendencies toward products or services. This data not only provides valuable information about consumer preferences to businesses but also offers opportunities for improving products and services to maintain a competitive edge in the market. For businesses, gaining a deep understanding of consumer sentiment regarding their products or services, and responding with corresponding improvements or adjustments, is of critical importance. However, traditional sentiment analysis methods have significant limitations when processing complex long texts. Sentiment analysis based on sentiment lexicons is time-consuming and labor-intensive, and single machine learning methods often fail to achieve satisfactory performance in feature extraction and semantic understanding. To address these challenges, this paper proposes a deep learning model for e-commerce review sentiment analysis based on a Convolutional Neural Network (CNN)-Bidirectional Long Short-Term Memory (BiLSTM) network architecture, incorporating the multi-head self-attention mechanism from the transformer model. This model aims to enhance the robustness and generalization of semantic relationships in long-distance dependencies and emotional information in the text, compensating for the shortcomings of traditional methods. Comparative analysis with several traditional machine learning feature extraction algorithms and mainstream deep learning methods shows that the proposed model outperforms these methods across various metrics and exhibits excellent portability and scalability.