Gesture Classification Based on System's Emotion and Strategy for Korean Dialogue Systems.

During the conversation, the speaker’s gestures can give more information to the listener and make it easier to understand the speaker’s utterance. Although appropriate system gestures based on the system strategy and emotion are crucial in a conversation between a person and a dialogue system, no Korean conversation corpora have system utterances annotated with both system strategies and emotions. In this paper, we pseudo-annotate system gestures on system utterances in the extended version of the Korean empathetic conversation corpus. Based on the pseudo-annotation, we introduce a system gesture list classification model on the basis of Korean-English T5 (KE-T5). We achieve an accuracy of 90.8% on the gesture list classification model.