This study leverages Google’s BERT pre-trained model, bert-base-chinese, to construct an innovative fusion model that integrates BERT with TextCNN. The model is developed using a dataset of smartphone reviews collected from a Chinese e-commerce platform, aiming to improve both prediction precision and model robustness in Chinese sentiment classification tasks. The dataset comprises 186,533 review texts, each annotated with a sentiment label. A fine-tuning process is first applied to the BERT pre-trained model to extract high-dimensional contextual representations, which are subsequently fed into the TextCNN model. Through convolution and pooling operations, TextCNN captures local textual features, and the fusion model undergoes a second fine-tuning process to improve sentiment polarity recognition. The evaluation findings indicate that the proposed BERT-TextCNN fusion model delivers better classification performance compared to traditional deep learning models including RNN, LSTM, and GRU, with the best test accuracy reaching 85.47%. However, while the fusion model improves overall performance, it also increases computational complexity, leading to higher training time and resource consumption. Additionally, the hyperparameters during fine-tuning remain fixed without extensive optimization, which limits further performance improvements. This study validates the potential and advantages of such a method in practical applications, providing a powerful tool for extracting user sentiment insights in the industry.