<p>Traditional student psychological intervention strategies often rely on static surveys and periodic assessments, which fail to capture the dynamic emotional fluctuations of students in real-time. To address this limitation, this paper proposes a novel Multimodal Sentiment Computing and Knowledge-Aware Recommendation System designed for proactive mental health monitoring in educational environments. The system is structured into a four-layer modular framework: Data Acquisition, Multimodal Feature Extraction, Sentiment Computing, and Intervention Decision. By integrating heterogeneous data streams—including textual logs, acoustic patterns, and facial expressions—the system utilizes a Multimodal Transformer (MulT) equipped with a Cross-Modal Attention mechanism to fuse features across distinct modalities and identify fine-grained psychological risks. Furthermore, a Knowledge-Graph-based Collaborative Filtering algorithm is implemented within the decision layer to map identified emotional states to personalized, evidence-based intervention strategies. Experimental results demonstrate the efficacy of this multimodal approach, achieving a 7.4% improvement in F1-score over the state-of-the-art single-modality BERT model. The suggested system utilizes a transformer-based approach for multimodal sentiment analysis through cross-modal fusion, which helps in combining textual, acoustic, and visual attributes for obtaining high accuracy with an F1-score of 0.906. Ablation studies and convergence analysis further validate the robustness of the system architecture and the necessity of cross-modal feature alignment. This research provides a scalable and privacy-preserving computational framework for higher education quality evaluation, offering a sophisticated tool for timely psychological support and academic engagement monitoring.</p>