A Multi-Scale Deep Learning Architecture for Psychological State Recognition and Early Risk Warning from Social Media Text

Social media has become an important channel for expressing emotional experiences and potential psychological distress, making automated psychological state recognition a key technical challenge for early risk warning systems. Psychological signals in text are distributed across multiple linguistic levels, ranging from character-level expressive variations to word-level semantics and sentence-level psychological structure, which limits the effectiveness of single-scale models. This paper proposes a multi-scale deep learning architecture for psychological state recognition from social media text. The approach integrates character-level and word-level representations, multi-scale convolutional modules for local semantic extraction, attention-based global semantic modeling, and cross-scale feature fusion. By jointly capturing fine-grained linguistic cues and global psychological context, the proposed model enhances the discriminative power of psychological representations. Experiments conducted on a multi-class mental health text dataset demonstrate that the proposed method consistently outperforms traditional machine learning models, conventional deep learning architectures, and attention-enhanced baselines in terms of accuracy, precision, recall, and F1-score. Furthermore, the model outputs are transformed into temporal risk signals, enabling the identification of weak, accumulating, and accelerating psychological risk patterns. The results indicate that multi-scale text modeling provides an effective technical solution for psychological state recognition and establishes a practical basis for the development of early psychological risk warning systems.