HSE-Net: Cognitively inspired asymmetric co-evolution of thematic structure and emotional dynamics in dialogue.

Modeling the interplay between macro-level structure and micro-level dynamics remains a fundamental challenge in neural sequence modeling, particularly in dialogue analysis. While cognitive theories establish that thematic evolution and emotional dynamics are intrinsically coupled-where topic appraisal drives emotion and affective states signal topic shifts-current neural dialogue models largely treat these dimensions in isolation. Consequently, existing architectures often oversimplify dialogue context into flat sequences or monolithic graphs, failing to capture the hierarchical dependencies essential for nuanced analysis. To bridge this gap, we propose the Hierarchical Synergistic Evolution Network (HSE-Net), a multi-scale architecture that mimics human cognitive processes by explicitly modeling the asymmetric co-evolution of thematic structure and emotional dynamics. At the core of HSE-Net is the Synergistic Co-evolution Block (SCEB), which implements an asymmetric feedback loop: it utilizes a content infusion mechanism to ground rapidly changing emotions in stable semantic contexts, while employing structural modulation to refine topic boundaries via affective cues. This design effectively disentangles task-specific representations while logically coupling their evolution, thereby mitigating the negative transfer common in naive multi-task learning. Furthermore, we facilitate this research by releasing the first high-quality topic segmentation annotations for standard emotion recognition benchmarks. Extensive experiments demonstrate that HSE-Net1 not only achieves state-of-the-art performance but also exhibits superior robustness against noise and improved latent space coherence, validating the efficacy of our cognitively inspired asymmetric co-evolution mechanism.