AI-Driven Intelligent Network Public Opinion Analysis Model for Universities with Dual-Dimensional Psychological and Emotional Features

A B S T R A C T The current public opinion analysis methods for colleges and universities have several drawbacks: single analysis dimension, separated sentiment and stance recognition, weak scene generalization ability, and absence of quantitative characterization of students’ psychological evolution rules, and these defects prevent intelligent online public opinion early warning and precise governance. This paper proposes a campus student public opinion analysis model (MDP-AM) integrating multi-task learning and two-dimensional psychological and sentiment features. RoBERTa serves as the underlying shared encoder, and the three subtasks including sentiment polarity analysis, stance tendency detection and sentiment-stance conflict detection are processed end-to-end simultaneously. In addition, emotion-knowledge enhanced prompt templates are adopted to boost the capture of fine-grained sentiment features. Meanwhile, LoRA technology that fine-tunes attention layer matrices is utilized to optimize the training process. Combined with the spiral of silence theory, this paper constructs two indicators: temporal correlation and cognitive correlation to quantitatively analyze public opinion evolution. Experimental results demonstrate that the model achieves substantially higher overall accuracy and F1 scores on the self-built campus public opinion dataset compared with mainstream baseline models. The two-dimensional psychological indicators can effectively implement early warning for high-risk public opinion and deliver a new technical solution for intelligent online public opinion governance in universities.