Interacting with a Sentimental Robot – Making Human Emotions tangible for a Social Robot via ChatGPT*

Recognizing and correctly assessing the emotions of a human conversation partner is – if successful – a milestone in social interaction between humans and social robots. The robot should recognize human emotions and take them into account in its reactions. It is also important that humans and robots assess the content of the dialog from an emotional perspective in the same way. But how can the emotional state of a person in a dialog with a robot be accessed? We examined this question more closely in our study. A Large Language Model (LLM) from OpenAI (ChatGPT) was used for conversations with a Pepper robot. We had the course of the dialog assessed once by the human interlocutor and once by the GPT model itself using sentiment analysis. In addition, the predominant emotion was named by both conversation partners after the dialog. A comparison of these evaluations provided an assessment of whether the human and the social robot arrived at the same results. We were also investigating whether the transmission of emotion recognition data had a noticeable influence on the tonality of the conversation. To do this, we used the robot’s emotion recognition capabilities to send cues to the GPT model about the current emotional state of the human at each turn of the conversation, so that the LLM could take this into account in the generation of the robot’s utterances. It was found that the predominant emotion of the human and the general mood of a conversation were interpreted by humans and the GPT model in largely the same way, whereby an existing emotion recognition made the robot’s assessment of the general mood deviate noticeably.