A Model for Perceiving Online Sentiment Risks in a Understandable Multimodal Context

With the growth of social media, online public opinion in higher vocational colleges exhibits multimodal characteristics. Traditional single-mode analysis methods struggle to accurately grasp the ideological trends of student groups. This paper proposes a Multimodal Social Media Sentiment Analysis Model (MSMSAM) based on large language models, which achieves deep fusion of multimodal data through image-to-text technology, and incorporates few-shot learning for risk prediction. On the experimental dataset of a specific higher vocational college, this model achieved an accuracy of 87.5% in sentiment classification tasks, marking a 15.7% improvement over baseline methods. The findings demonstrate that multimodal fusion strategies can effectively enhance the accuracy of online public opinion analysis, providing a new technical pathway for mental health early warning in colleges.