Subjective Evaluation of Dialogues with Dyadic and Triadic Interactions Using a ChatGPT-Based Dialogue System

We investigated subjective evaluation of dialogues with dyadic and triadic interactions by a questionnaire. This system uses ChatGPT-based conversation with a facial expression function. In terms of human developmental stages, communication on triadic interaction is more advanced than on dyadic interaction. However, dialogue on triadic interaction was not evaluated as better necessarily. On dyadic interaction, most subjects evaluated this system as a good partner. On triadic interaction, half of the subjects evaluated this system as a very good partner who listened carefully, while another half evaluated as a poor partner because they couldn't communicate well. We estimate the reason of this negative evaluation is that the dialogue structure, which was designed to resemble interpersonal communication, required communication effort from the subjects, and they could find that a burden.