Behavioural Observations as Objective Measures of Trust in Child-Robot Interaction: Mutual Gaze

Sammanfattning: Given the subjective nature of trust as a phenomenon and its unified multifaceted contributions for every individual context,the development of a computational model of trust proves to be a difficult endeavour. In this study, we investigate mutual gaze as a behavioural measure of social trust and liking in child-robot interaction. Developing on a prior user study involving 52 children interacting with a robot with variable human-likeness and lexical alignment in two interaction contexts (task-based and dialogue-based), we investigate the effects of human-likeness and lexical alignment on mutual gaze, associations and correlations between metrics assessing social trust and liking, and the development of mutual gaze as an objective measure of social trust and liking. We achieve this through several statistical analyses between the percent of mutual gaze in each interaction, human-likeness, lexical alignment, scores from social trust and liking metrics, self-disclosure content, age, and time. The main findings of our study support the use of mutual gaze as an objective measure for liking, but there is still not sufficient evidence to supportthe use of mutual gaze as an objective measure to identify and capture social trust as a whole. Furthermore, we found that human-likeness and lexical alignment do not significantly affect mutual gaze in an interaction, but the interaction context does. Moreover, it seems that age plays a role in the amount of mutual gaze in an interaction, where older participants engage in less mutual gaze compared to the younger participants. Alongside this, the amount of mutual gaze the participant engages in is stable across periods when they are not interacting with the robot, changing more towards the first half of the first interaction and the second half of the second interaction. Based on the study, our findings suggest using different objective behavioural measures for social trust compared to its related concepts such as liking. Also, our results have found that there may be other constructs intertwined with liking, such as attention and interest, which may need to be addressed with separate metrics.