Algorithm Design for User Emotion Perception and Psychological State Prediction in Large Language Model Interaction

Abstract

With the continuous evolution of intelligent communication systems and distributed information processing technologies, real-time perception of user states has become an important component of adaptive human–machine interaction. This study proposes a dual-task parallel algorithm for instantaneous emotion perception and deep psychological state prediction in large language model interaction under pure text communication. The framework jointly models lexical features, syntactic complexity, interaction rhythm, topic coherence, and language habits to achieve synchronized short-term emotion tracking and long-term psychological state inference, while introducing an uncertainty estimation mechanism for confidence-aware decision making and interaction avoidance. Experimental results demonstrate that the proposed method achieves a macro F1 score of 0.85 for instantaneous emotion classification and a Pearson product-moment correlation coefficient of 0.76 for deep psychological state prediction, outperforming representative baseline approaches. Furthermore, confidence-driven interaction avoidance effectively suppresses high-risk misjudgments and improves overall interaction reliability. By integrating multi-level semantic perception with uncertainty-aware inference, the proposed framework provides an efficient solution for adaptive information processing and intelligent decision support, offering valuable methodological references for real-time signal interpretation, human–machine communication, and intelligent perception systems in Electromagnetic Waves, Antennas and Propagation related applications.