EmoShiftNet: a shift-aware multi-task learning framework with fusion strategies for emotion recognition in multi-party conversations

Introduction Emotion Recognition in Conversations (ERC) is vital for applications such as mental health monitoring, virtual assistants, and human–computer interaction. However, existing ERC models often neglect emotion shifts—transitions between emotional states across dialogue turns in multi-party conversations (MPCs). These shifts are subtle, context-dependent, and complicated by class imbalance in datasets such as the Multimodal EmotionLines Dataset (MELD). Methods To address this, we propose EmoShiftNet, a shift-aware multi-task learning (MTL) framework that jointly performs emotion classification and emotion shift detection. The model integrates multimodal features, including contextualized text embeddings from BERT, acoustic features (Mel-Frequency Cepstral Coefficients, pitch, loudness), and temporal cues (pause duration, speaker overlap, utterance length). Emotion shift detection is incorporated as an auxiliary task via a composite loss function combining focal loss, binary cross-entropy, and triplet margin loss. Results Evaluations on the MELD dataset demonstrate that EmoShiftNet achieves higher overall F1-scores than both traditional and graph-based ERC models. In addition, the framework improves the recognition of minority emotions under imbalanced conditions, confirming the effectiveness of incorporating shift supervision and multimodal fusion. Discussion These findings highlight the importance of modeling emotional transitions in ERC. By leveraging multi-task learning with explicit shift detection, EmoShiftNet enhances contextual awareness and offers more robust performance for multi-party conversational emotion recognition.