Emotion Recognition

Emotion recognition studies how to infer what people feel from speech, text, facial and bodily behavior, and physiological signals, for uses such as communication support and human-computer interaction. Common approaches include neural classifiers and continuous valence-arousal regression, multimodal fusion, conversational and scene-context modeling, and newer foundation-model or language-model methods; benchmarks include IEMOCAP, GoEmotions, MELD, Aff-Wild2, DEAP, and SEED. Work increasingly models uncertainty and mixed emotions with annotator distributions, soft labels, and temporal trajectories rather than treating each example as a single fixed label. A central challenge is that emotion is subjective and culturally and contextually variable, while models often struggle to generalize across people, domains, and real-world settings without sacrificing privacy or interpretability.