Research on job burnout prediction model based on emotion-semantic dual-channel representation and reinforcement learning

Job burnout has become an ‘occupational epidemic’ that threatens the health of the global labor force. Traditional questionnaires and manual counseling are slow to respond and narrow in coverage. The existing AI methods are disjointed between multimodal representation and decision optimization, lack of dynamic alignment of emotion-semantic dual-channel features, and conflict between self-supervised pre-training and reinforcement learning goals, resulting in poor generalization and homogeneous intervention. To address these challenges, this study proposes the EmoSem-IRL framework, which constructs a RoBERTa–NRC-based emotional channel and a ConceptNet-enhanced semantic channel, dynamically integrated through interactive gating (a mechanism that adaptively balances emotion and semantic information). By introducing KL divergence loss constraints (used to align the feature distributions of the two channels), PPO reinforcement learning is employed to generate 12 types of personalized interventions on the fused features, forming a self-supervised–reinforcement closed-loop optimization (an end-to-end learning cycle combining representation and decision optimization). The framework’s novelty lies in bridging emotion and semantics via external knowledge graphs and enforcing stable dual-channel alignment through KL-based optimization. The cross-industry test shows that the average absolute error of the medical industry is reduced to 0.19, and the F1 of financial, IT and medical high-risk groups is ≥ 0.81. The dual-channel KL divergence decreased from 0.56 to 0.18, a decrease of 67.9%. The average adoption rate of 12 types of interventions was 67.2%. After 4 weeks of follow-up, 65.6% of users’ burnout decreased by more than 30%. EmoSem-IRL realizes the end-to-end coordination of knowledge-enhanced dual-channel representation and closed-loop enhanced decision-making, which significantly improves the prediction accuracy of burnout and the personalization of intervention, and provides a new paradigm for generalized and sustainable intelligent mental health management. Emotion–Semantic Dual Channel Innovation: The RoBERTa–NRC emotional channel runs in parallel with the ConceptNet-enhanced semantic channel, enabling explicit modeling of emotional–semantic inconsistency in textual expression. Dynamic Fusion and Alignment Mechanism: An interactive gating mechanism with bidirectional context weighting integrates dual-channel features, while KL divergence–based alignment and orthogonal projection achieve conflict perception and soft correction within a unified optimization loop. Closed-Loop Self-Supervised Reinforcement Learning: A self-supervised pretraining strategy is grafted onto PPO reinforcement learning to generate 12 types of personalized interventions driven by multi-objective rewards, achieving fine-grained personalization and cross-industry generalization.