CEREBRAL: A Neurosymbolic Framework for Multimodal Emotion Recognition with Psychological Constraints and Metacognitive Reasoning

Multimodal emotion recognition remains difficult due to cross-modal dependencies, temporal dynamics, and the need for psychologically consistent, interpretable outputs. We introduce CEREBRAL, a neurosymbolic architecture that fuses neural multimodal processing with symbolic reasoning and metacognitive control. It uses Answer Set Programming for logical inference, encodes the Hourglass of Emotions as four-dimensional affective constraints with dynamic polarity normalization and sentic vectors, and incorporates Neural Turing Machines for episodic memory and Graph Neural Networks for temporal consistency. CEREBRAL processes fine-grained emotions through cross-modal attention, dynamic memory, and metacognitive strategy selection with uncertainty estimation. We evaluate CEREBRAL across multiple benchmark datasets, where it consistently outperforms neural-only baselines while preserving high symbolic reasoning accuracy with complete logical proof generation. Statistical significance testing confirms these improvements with robust performance under noise conditions and cross-dataset generalization. The symbolic reasoning component demonstrates practical efficiency and generates human-interpretable explanations through Hourglass dimensional analysis. This work contributes a psychologically grounded approach to emotion recognition that balances neural learning with symbolic constraints, offering interpretability alongside performance gains. The framework’s explicit reasoning traces, four-dimensional affective representation, and calibrated uncertainty estimates address key requirements for deploying emotion-aware AI in clinical settings, human-computer interaction, and affective computing applications where transparency and reliability are essential.