Knowledge Augmented Ambiguity Aware Learning for Multi-Label Emotion Detection

We propose KAME, a Knowledge-Augmented Ambiguity-Aware Learning framework for robust multi-label emotion detection. KAME enriches transformer-based representations with zero-shot semantic priors obtained from an external natural language inference model and introduces entropy-guided ambiguity weighting to reduce the influence of highly uncertain training instances. The framework further incorporates R-Drop consistency regularization, exponential moving averaging, and multi-seed/multi-backbone probability ensembling with development-set threshold optimization. Experiments on English and Chinese multi-label emotion datasets demonstrate strong and balanced performance. KAME achieves a Macro-F1 of 0.650 and Hamming loss of 0.117 on English, while obtaining the best Micro-F1 of 0.673, Macro-F1 of 0.616, and Hamming loss of 0.136 on Chinese among the compared methods. Ablation results further confirm the importance of semantic priors, consistency regularization, EMA, and ensemble aggregation. Overall, KAME provides an effective framework for integrating external semantic knowledge, uncertainty-aware learning, and robust prediction for multi-label emotion recognition.