Building and Assessing an Italian Textual Dataset for Emotion Recognition in Human‐Robot Interactions

Abstract

In this study, we illustrate an ongoing work regarding building an Italian textual dataset for emotion recognition for HRI. The idea is to build a dataset with a welldefined methodology based on creating ad-hoc dialogues from scratch. Once that the criteria had been defined, we used ChatGPT to help us generate dialogues. Human experts in psychology have revised each dialogue. In particular, we analyzed the generated dialogues to observe the balance of the dataset under different parameters. During the analysis, we calculated the distribution of context types, gender, consistency between context and emotion, and interaction quality. With "quality" we mean the adherence of text to the desired manifestation of emotions. After the analysis, the dialogues were modified to bring out specific emotions in specific contexts. Significant results emerged that allowed us to reorient the generation of subsequent dialogues. This preliminary study allowed us to draw lines to guide subsequent and more substantial dataset creation in order to achieve increasingly realistic interactions in HRI scenarios.