Adversarial Multi-feature Fusion for Sentiment Analysis

Existing text sentiment classification methods suffer from several limitations: they struggle to capture global semantic information, fail to fully utilize positional and relational features of words, and exhibit inadequate handling of long-range semantic dependencies or emotional transitions. Moreover, the suboptimal fusion of features often leads to unsatisfactory classification accuracy. To address these issues, this paper proposes a multichannel sentiment analysis model. First, BERT is employed for data preprocessing. Next, Graph Convolutional Networks (GCN) are utilized to extract part-of-speech and syntactic features, while Convolutional Neural Networks (CNN) capture positional and local semantic features. Additionally, Bidirectional Simple Recurrent Units (Bi-SRU) are applied to model contextual dependencies, enabling deep exploration of textual features. Finally, a Generative Adversarial Network (GAN) is introduced to fuse these features, enhancing their complementarity. Ablation experiments validate the effectiveness of each module, and comparative results on public datasets demonstrate that the proposed model outperforms existing baseline methods, showcasing its superior performance.