Multilabel Emotion Recognition Using a Label-Query Dynamic Dependency Network

A label-query decoding module is introduced to extract emotion-specific representations from contextualized token features, enabling more discriminative modeling of individual emotions. Furthermore, a dual-branch dependency modeling framework is proposed to capture both static corpus-level label relationships and dynamic instance-level label interactions, allowing the model to better exploit inter-emotion dependencies. Finally, a joint optimization strategy is adopted by integrating classification, ranking-aware, and calibration-aware objectives, thereby improving multilabel decision quality and enhancing the prediction of infrequent or rare emotion labels.