Emotion Recognition in Conversations
Overview
Emotion Recognition in Conversations aims to identify how people feel from dialogue, accounting for both the words they use and how emotions shift across turns and speakers. Common approaches combine text, speech prosody, and facial or visual features using transformers, speaker-aware graphs, recurrent or state-space models, and adaptive fusion; some also track valence and arousal or represent uncertainty instead of assigning a single fixed label. Research commonly evaluates on benchmarks such as IEMOCAP and MELD, while newer work examines longer-range memory, implicit affect, streaming speech, and emotion-cause relations. A central challenge is that labels can be ambiguous and modalities unreliable or conflicting, so high benchmark scores do not always guarantee robust interpretation across speakers, settings, and cultures.
Latest papers
61 papers4 papers in the last four weeks, down 64% on the four weeks before. 0.0% of all new papers.