Aspect-level sentiment analysis based on fast gradient method and semantic fusion

Aspect-level sentiment analysis is difficult to effectively integrate global context and local details due to the complexity of spoken expressions, which limits the extraction of fine-grained features. This paper proposes a framework that combines adversarial training Fast Gradient Method (FGM) and Semantic Fusion Attention-Convolution Module (SFAC). FGM enhances robustness through perturbations, and SFAC combines self-attention with convolutional networks to dynamically integrate global semantic and local syntactic features to optimize emotional interaction. Experiments show that the model significantly improves classification performance on four public datasets. Ablation studies verify the key role of FGM and SFAC in enhancing robustness and capturing emotional patterns, highlighting the effectiveness of the fusion of semantic and syntactic modeling.