EAAN-ERC: Expert Adaptive Agreement Network for Emotion Recognition in ConversationExpert Adaptive Agreement Network for Emotion Recognition in Conversation • April 1st, 2024
Existing studies for emotion recognition in con- versation (ERC) focus on modeling conversa- tional context, however, they overlook the in- fluence of diverse human evaluator panels on the emotional annotations of datasets. We ob- served in an existing ERC dataset that differ- ent evaluator panels for assessed utterances in conversations impact the final emotional eval- uation results due to the subjective nature of each evaluator’s perception and interpretation of emotion. To address this issue, we pro- pose a novel Expert Adaptive Agreement Net- work for Emotion Recognition in Conversation (EAAN-ERC), a method designed to imitate the evaluation and annotating process of emo- tions by diverse evaluator panels. Specifically, we first mimic experienced evaluators by set- ting up multiple expert models. Subsequently, we emulate diverse evaluator panels by adap- tively mixing expert models matched with spec- ified evaluator panels. Furthermore, we simu- late the evaluator panels’ emotiona
