About

Virtual agents are used in a number of different outcome-based contexts such as physical and mental health, skill-based training, as well as classroom learning and pedagogy. Virtual agents in such applications are largely designed so that they project positive attitude and feedback towards the human participant. Human-human interactions, however, are certainly not exclusively positive in valence. For example, teachers and educators engage in both positive and negative feedback strategies for pedagogical outcomes. While the distinct effects of positive and negative feedback on learning are well established, few studies have attempted to examine the effects of negative feedback across different combinations of instructor's gender and proxemics-based physical behavior. This study explores this very question with a 2 (instructor gender)*2 (proxemic behavior) between subject design. In this experiment, participants (N=63) actively engage in a learning task with a male/female virtual instructor that provides negative feedback while either standing stationary or while physically approaching the participant. Based on the different deliveries of the negative feedback, the study aimed to identify the sources of variations in participant reactions to the negative feedback, namely patterns of attribution and both behavioral and physiological measurements of emotions. The results indicate that participants attribute greater self-blame (internal attribution) for their purported poor performance when interacting with the female virtual instructor than when interacting with the male virtual instructor. Participants also generally exhibited greater positive affect in response to female virtual professors than male virtual professors. These results are highly relevant both to the design of virtual agents as well as to adding to our understanding of the role of gender and behavior in human-human, non-peer interaction.

Method

The inability to satisfy the person of power in this interaction is suggested to lead to feelings of helplessness and lack of motivation, key components of negative feedback. Using this scenario as a template, we created a virtual environment in which individuals are assigned to a task that they are incapable of completing to the satisfaction of an instructor.

Specifically, the interactive virtual environment here simulated an acting class scenario. One of the virtual characters was designed to be the instructor in the scene. Each participant and other non-player characters (NPC) were students who were asked to rehearse 'Romeo and Juliet: Act 3, Scene 3'.

Each time the participant finished reading a line, the virtual instructor provided negative feedback in a number of ways including harsh language, negative non-verbals, encroaching on personal space, and ridiculing the participants' performance. Although the negative feedback from virtual instructor were scripted and identical for all participants, participants were told the feedback was tailored based on their performance and they should follow the instructor's directions to the best of their ability. At the conclusion of the experiment, participants were debriefed about the scripted and non-authentic nature of the "feedback".

To invoke negative affect, the system utilized social interaction and an impossible task framework as mechanisms. All feedback given by the virtual instructor, regardless of actual performance, were designed to be negative and variable in nature.

An example of a trial

The Revised Causal Dimension Scale (CDSII) (Mcauley et al. 1992) was used to measure assignment of causal attributions after the conclusion of the experiment. The CDSII consists of four individual dimensions, "locus of causality" (internality), "stability", "personal control", and "external control". Responses are made on 9-point semantic differential scale with anchoring statements at either end of the scale.

In addition to the original items, items tailored for this experiment were included as slight modifications to the existing items. For instance, "Professor can regulate : professor cannot regulate", "That reflects an aspect of yourself : reflects an aspect of the professor" and "Something about you : something about the professor".

The scale used in the experiment is listed below.

Project Overview

Upon completion of the experiment, participants were also asked to respond to an ad-hoc questionnaire on a 5-point Likert scale designed to comprehensively evaluate the system's performance.

The ad-hoc questions are listed below

  • 1. To what extent were you trying hard to perform the task?
  • 2. Was the feedback you got on your performance helpful?
  • 3. How accurate was the feedback on your performance?
  • 4. How likable was the virtual professor?
  • 5. To what extent were you concerned about the professor’s evaluation of your performance on the task?
  • 6. Do you think the professor's reactions were due to your underperformance?
  • 7. Do you think the professor's reactions were due to him/her having a bad day?
  • 8. Do you think the professor's reactions were due to his/her personality?
  • 9. Assume you have the opportunity to do an advanced degree with a world famous professor who instructed like this, how likely is it that you would take this opportunity?
  • 10. Would you consider taking an acting class from a different professor?
  • 11. If a professor performed like this, would you report him/her to the university?
  • 12. If a professor performed like this, would you report him/her to your family?
  • 13. If a professor performed like this, would you report him/her to friends?
  • 14. If a professor performed like this, would you report him/her to fellow students?

Result

Publication

Acknowledgment

We thank Stephen J. Read for his help in selecting psychological measures for this study and his suggestions on modifications of the CDSII scale.

This research was supported by the Air Force Office of Scientific Research under grant FA9550-14-1-0364. Statements and opinions expressed do not necessarily reflect the position or the policy of the United States Government, and no official endorsement should be inferred.