Image–Text Multimodal Sentiment Analysis Algorithm Based on Curriculum Learning and Attention Mechanisms

With the rapid development of mobile internet technology, the explosive growth of image–text multimodal data generated by social networking platforms has provided rich practical scenarios and theoretical research value for multimodal sentiment analysis. However, existing methods generally suffer from inefficient modal interaction and imperfect sentiment aggregation mechanisms, particularly an over-reliance on visual modalities, leading to an imbalance in cross-modal semantic correlation modeling. To address these issues, this paper proposes a sentiment analysis algorithm for image–text modalities based on curriculum learning and attention mechanisms. The algorithm introduces the concept of curriculum learning, fully considering the negative impact of irrelevant images in image–text data on overall sentiment analysis, effectively suppressing interference from irrelevant visual information without requiring manual data cleaning. Meanwhile, the algorithm designs a dual-stage attention architecture—first capturing cross-modal correlation features via cross-modal attention, then introducing an attention bottleneck strategy to compress redundant information flow, achieving efficient feature fusion by constraining intra-modal attention dimensions. Finally, extensive experiments were conducted on two public datasets, demonstrating that the proposed method outperforms existing approaches in sentiment prediction performance.