Strategies for Controlling the Conversation Dynamics in Multi-Party Human-Robot Interaction

This article tackles the research question of whether it is possible to control conversation dynamics in a multi-party scenario using easily implementable solutions on off-the-shelf robotic platforms. To this end, we expanded upon our previously developed cloud robotic architecture by incorporating policies aimed at managing conversation dynamics through selective addressing of individuals, with the ultimate goal of balancing or unbalancing users’ participation or making subgroups of participants interact. Specifically, we computed the dominance of each speaker as a weighted sum of their speaking time and the number of words spoken within a moving window and used the Louvain algorithm to partition speakers into a set of non-overlapping communities. We then implemented six control policies, which were applied by the robot. Two of them, named BH and BS, aim to reduce dominance error (i.e., the difference in dominance between the most and least dominant speakers—both policies give the floor to the less dominant speaker). Two other policies, UH and US, are designed to increase the dominance error (both give the floor to the most dominant speaker). Finally, CH and CS aim to reduce the community error (i.e., the difference between the actual number of detected subgroups among speakers and the ideal target of a single group to which all speakers belong). Policies BH, UH, and CH (with “H” standing for “hard”) do not allow any exceptions to the policy rules, while BS, US, and CS (with “S” for “soft”) permit exceptions. To test the impact of these policies, we conducted a between-subjects study (N = 300) involving middle school students engaging in dialogue with a humanoid robot acting as a moderator. The study compared five conditions: in four of them, the robot used information gathered during the conversation to decide which speaker to address, applying one of the control policies—BH, BS, CH, or CS. The policies UH and US were excluded, as having a robot consistently give the floor to the most dominant child may raise ethical concerns. In the fifth condition, a baseline neutral policy (N) was applied, in which the robot did not explicitly address any speaker. The results imply that a robot using the proper control policies can influence conversation dynamics to keep both dominance error and community error significantly lower than those of a robot using the baseline policy, leading to more balanced participation and a reduction in the number of subgroups. Indeed, statistically significant differences have been found between the five policies considered in the dominance and community errors. However, no statistically significant differences in user experience—as measured by three scales of the validated SASSI questionnaire—were found when the robot used one of the control policies, as compared to the baseline, suggesting that participants are not negatively impacted by the robot’s attempt to control the conversation.