Sentiment analysis for tourism reviews based on dual-stream graph attention fusion network

Emotional evaluations can serve as indicators to guide the improvement of online travel service quality. This paper proposes a model based on a dual-stream graph attention fusion network. To prevent the introduction of noise unrelated to the subject matter, the network encodes the aspect texts and integrates them into the image-text embeddings through cross-attention mechanisms. Furthermore, the network incorporates a dual-stream interactive masking graph neural network to effectively align and fuse the deep features of multimodal image and text data. This allows the image data and textual data to provide rich supplementary information for each other, thereby enhancing the model's accuracy in sentiment analysis. In simulation comparisons with various state-of-the-art methods on real-world multimodal travel datasets, the results demonstrate that this approach has reduced the MAE and RMSE metrics by 19.86% and 5.26%, respectively, compared to the best-performing baseline model.