Conceptual
Browse Instruct Curate

Turn-Priority Dialogue-Feature Attention for Emotion Recognition in Conversation

TED is an emotion-recognition-in-conversation method that injects dialogue features -- turn position and speaker identity -- into multi-head self-attention. A Turn-Based Encoding builds a per-turn vector from a pretrained language model; a priority factor reweights the attention scores by inter-turn distance and speaker match; and a dialogue layer runs self-attention over the turn sequence to label each turn's emotion. It targets long conversations with many turns, reaching state-of-the-art on IEMOCAP.