Application of BERT-EmoShiftNet fusion framework in psychological stress dialogue

Psychological stress dialogue recognition still lacks sufficient semantic understanding in multimodal information fusion and emotional dynamic modelling. It is difficult to achieve high-precision and robust stress state recognition in complex dialogue scenarios. Combining the advantages of the pre-trained language model BERT (Bidirectional Encoder Representations from Transformers) in semantic understanding and the capabilities of EmoShiftNet (Emotion Shift Network) in emotion transfer modelling, contextual semantic coding is realised through a multi-level Transformer, and a cross-modal Transformer is used to align heterogeneous features such as text, speech, and vision. In terms of multimodal signal representation, the BERT encoder captures fine-grained changes in language semantics, the convolutional-recurrent network extracts speech time–frequency features, the convolutional neural network and optical flow method obtain facial expression dynamics, and finally, realises unified representation through shared embedding space and cross-modal contrast learning. On the five-class psychological stress dialogue task, text semantic features contribute 42.7% of the total discriminative information, and the full multimodal fusion model achieves an overall accuracy of 85.0% (95% CI (Confidence Interval): [83.2%, 86.8%]) on the held-out test set. In the emotion transition detection task, the model achieves 76.2% transfer prediction accuracy with an average detection latency of 13.4 s. Under low signal-to-noise ratio conditions, the speech modality maintains stable performance with an error rate of 4.8% across 59 test samples. The highest accuracy rate of the fusion model under full-modal input reaches 97% on our self-constructed psychological stress dialogue dataset, which is divided strictly at the dialogue level (6391 training, 1484 validations, 2227 test samples) to avoid data leakage. This result significantly outperforms single-modal baselines and four state-of-the-art pre-trained language model baselines.