Hierarchical Attention Mechanisms for Multimodal Sentiment Analysis in User-Generated Social Media Content

In multimodal sentiment analysis, effective extraction of fused multimodal features and the use of attention mechanism to improve analysis accuracy are key issues. The research constructs multi-source heterogeneous data collection architecture to collect user-generated content containing text, images, and videos from social media, short videos, news media, and other platforms. This study proposes a multimodal sentiment analysis model based on bidirectional long short-term memory network and hierarchical attention mechanism. It introduces the sentiment attention capsule structure and combines it with a sentiment loss function to guide the generation of specific sentiment responses. The experiments are based on a multimodal emotion intensity dataset and a multimodal dialog action dataset, which are compared with multiple models. The results showed that the model had an accuracy of 90.285%, precision of 90.442%, recall of 89.517%, F1-score of 89.980%, AUC value of 0.935, and training time of 9.56 hours. Moreover, the overall performance was significantly better with 64.887 ms inference speed, 2.487 GB memory usage, and 94.209% training stability. The study offers a fresh approach to raising sentiment analysis accuracy, and the model's attention mechanism and data fusion architecture are novel. It has application value for intelligent dialogue, emotion monitoring and other scenarios, and provides reference for model optimization and technology landing in this field.