A Multimodal Aspect-Level Sentiment Analysis Model based on Syntactic-Semantic Perception

Current multimodal aspect-level sentiment analysis models often face two primary challenges. Firstly, there is insufficient integration of syntactic and semantic information during syntactic analysis. Secondly, there is excessive noise during the fusion of different modalities. This paper proposes a multimodal aspect-level sentiment analysis model based on syntactic-semantic perception to address these issues. First, the model extracts contextualised textual representations using a pre-trained language model (BERT), capturing syntactic information by constructing dependency and phrase structure trees for each sentence. It then uses Graph Convolutional Networks (GCNs) to learn contextual textual features, helping to reduce errors when identifying aspects and their corresponding opinions. The feature fusion module introduces a learnable gating mechanism that dynamically adjusts the weights of different features, enabling the model to adaptively determine each feature’s contribution to sentiment prediction. Finally, the fused features are fed into a softmax layer for sentiment classification. Experimental results demonstrate that the proposed model achieves accuracy rates of 72.8% and 71.1% on the Twitter15 and Twitter17 benchmark datasets, respectively, achieving a significant improvement on baseline methods.