A semantic fusion depression recognition model integrating textual part-of-speech and sequential features

Depression is a mental disorder that significantly impairs patients’ emotional and functional well-being status. With the breakthrough development of natural language processing technology, text-based depression recognition has emerged as a promising non-invasive diagnostic approach. Nevertheless, traditional models often suffer from limited feature representation and contextual ambiguity, particularly in Chinese. This study proposes a novel semantic fusion depression recognition model (SFDRM) was proposed by integrating part-of-speech (POS) features with sequential features of text to address these challenges. First of all, an improved TF-IDF algorithm was developed by incorporating the sentiment dictionary with semantic rules, thereafter extracting sentiment-oriented POS based. Second, global semantic representations were extracted using BERT, and local contextual features are captured via BiLSTM. Finally, a self-attention mechanism was introduced to dynamically fuse these multi-granular features, enabling the model to emphasize salient information. Experiments conducted on two publicly available Chinese depression corpora (EATD and CMDC) showed that SFDDM has higher accuracy (0.86) and F1 score (0.80) compared to traditional and single feature models. This indicated that the model effectively addressed the ambiguity problem in Chinese sentiment analysis, and effectively improved the generalization ability and robustness of the model. Therefore, the current SFDRM might provide a more effective and universal solution for intelligent depression screening based on Chinese language patterns, and lays the foundation for the integration of acoustic features in multimodal mental health assessment in the future.