How (not) to Evaluate Computational Empathy: Testing the Assumptions of the Evaluation Methods in a Use-Case

An increasing amount of successful contributions were made in the past two decades to model empathic behavior in artificial agents to enhance the interaction between humans and artificial agents in variety of contexts. However, this relatively novel research area still did not develop standard methods for evaluation. In this paper, we use an embodied conversational agent (ECA) with real time multi-modal interaction capabilities, to test the assumptions of the proposed evaluation method of computational empathy in a patient intake scenario of counselling services in university environments. Our findings show that our system was able to show more empathy compared to its human counterpart and that the proposed evaluation method was supported by the user study in determining the changes in users' perception in empathic behavior by relying on feature-level metrics and other factors that are not directly related to empathy, which are shown to have a significant influence on the perception of empathic behavior. However, our detailed analyses showed poor reliability of the empathy questionnaire used for the system evaluation, suggesting the need for a metric that is tailored for artificial agents.