Abstract
Traditional statistical learning-based sentiment analysis methods often struggle to effectively handle text relevance and temporality. To overcome these limitations, this paper proposes a novel approach integrating Latent Dirichlet Allocation (LDA), Shuffle-enhanced Real-Valued Non-Volume Preserving (RealNVP), a double-layer bidirectional improved Long Short-Term Memory (DBiLSTM) network, and a multi-head self-attention mechanism for sentiment analysis. LDA is employed to extract latent topics within comment texts, revealing text relevance and providing fine-grained user feedback. Shuffle enhancement is applied to RealNVP to effectively model the distribution of text topic features, enhancing performance while avoiding excessive complexity in model structure and computational overhead. The double-layer bidirectional improved LSTM, through the coupling of forget and input gates, captures the dynamic temporal changes in sentiment with greater flexibility. The multi-head self-attention mechanism enhances the model's ability to select and focus on key information, thereby more accurately reflecting user experiences. Experimental results on both Chinese and English online comment datasets demonstrate that the proposed integrated model achieves improved topic coherence compared to traditional LDA models, effectively mitigating overfitting. Furthermore, the model outperforms single models and other baselines in sentiment classification tasks, as evidenced by superior accuracy and F1 scores. These results underscore the model's effectiveness for both Chinese and English sentiment analysis in the context of online comments.