HCC: Medium: Agent-Facilitated, Video-Mediated Multiparty Interactions in Support Groups

支持小组帮助人们向有相似经历的人学习；众所周知，它们能有效地减轻由负面生活事件引起的压力。随着互联网的广泛使用，支持小组也扩展到视频会议形式。现场和远程支持小组由具有广泛背景和资格的辅导员领导。不幸的是，这样的引导者经常遭受倦怠，导致支持小组关闭。因此，当人工辅助器不可用时，自动化辅助器提供了一种维护支持小组的方法。这个项目的主要目的是：(i)识别和评估一个有效的自主小组促进者的特征；（ii）研究和开发用于测量视频介导的多方互动中的个人参与度和群体凝聚力的计算方法；（iii）开发和评估一个能够通过计算手段最大化群体凝聚力的自主群体促进者。为了实现这些目标，本项目通过Zoom或类似平台为远程支持团体构建并研究了一种以社交辅助机器人形式的自主代理促进者。团体中的人际联系和联盟使支持团体更有效。因此，该项目将使机器人引导者能够选择促进策略，增加小组成员的参与度和连通性。这项研究推动了人工智能技术的发展，以理解人机交互，并有助于开发能够扩大获得精神卫生支持的技术。项目活动将包括为当地市中心K-12学生举办的年度外展会议，展示自动化促进者并讨论压力管理，以教育STEM和心理健康。该项目还将通过K-12外展活动，以及每年从系统服务不足的群体中培训和指导五名本科研究人员，扩大对计算机的参与。该项目推进了能够与多个用户进行视频交互的社交交互代理和机器人的最新技术。该研究结合了表达机器人和代理实施体的研究，自主对话促进算法，以及新型促进策略的用户参与。为此，该项目将首先使用人类驱动的代理，通过奥兹巫师（WoZ）策略，设计代理的行动空间（包括口头和非口头行为），以调节支持小组。WoZ研究还将测试一个假设，即在吸引用户和向群体参与者投射能力方面，一个具身代理促进者与一个人类促进者一样有效。在对WoZ研究期间记录的数据进行编码后，将训练多模态机器学习模型，以自动识别参与和对话阶段和行为。通过网络分析，团队凝聚力将基于二元参与和个人反应进行评估。研究团队最终将建立一个自主促进者，利用强化学习模型，优化提高团队凝聚力。在个人参与和小组凝聚力方面，自主促进者将与第二代理进行评估，第二代理优化了平等进入会话层的机会，并通过会后问卷评估。这项工作将建立自动化小组促进技术，在人工促进人员缺席或短缺的情况下，帮助弥合提供支持小组的差距。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Support groups help people to learn from others who share similar experiences; they are known to be effective in reducing stress caused by negative life events. With broader access to the Internet, support groups have also expanded into video conferencing format. In-person and remote support groups are led by facilitators with wide ranging backgrounds and qualifications. Unfortunately, such facilitators often suffer from burnout, leading to support group closure. Hence, automated facilitators offer a way for maintaining support groups when human facilitators are unavailable. The main aims of this project are (i) to identify and evaluate the characteristics of an effective autonomous group facilitator; (ii) to study and develop computational methods for measuring individual engagement and group cohesion in video-mediated multiparty interaction; and (iii) to develop and evaluate an autonomous group facilitator that can maximize group cohesion through computational means. To achieve these aims, this project builds and studies an autonomous agent facilitator in the form of a socially assistive robot for remote support groups via Zoom or a similar platform. The interpersonal connectedness and alliances in a group make a support group more effective. Therefore, the project will enable the robot facilitator to choose the facilitation strategy that increases group members’ participation and connectedness. This research advances AI technologies for understanding human-robot interaction and contributes to the development of technologies that can broaden access to mental health support. The project activities will include annual outreach sessions for local inner-city K-12 students demonstrating the automated facilitator and discussing stress management, to educate about STEM and mental health. This project will also broaden participation in computing through the K-12 outreach activities and through training and mentoring five undergraduate researchers per year from systematically underserved groups.This project advances the state-of-the-art in socially interactive agents and robots capable of interacting with multiple users, in video-mediated interaction. The research incorporates the study of expressive robot and agent embodiment, algorithm for autonomous conversation facilitation, and user engagement for novel facilitation strategies. To this end, the project will first use a human-driven agent, through a Wizard-of-Oz (WoZ) strategy, to design the agent’s action space (both verbal and nonverbal behaviors) necessary for moderating a support group. The WoZ study will also test the hypothesis that an embodied agent facilitator is as effective as a human facilitator in engaging users and projecting competence to group participants. After coding the data recorded during the WoZ study, multimodal machine learning models will be trained for automatic recognition of engagement and conversational stages and acts. Group cohesion will be assessed based on dyadic engagement and individual responses, through network analysis. The research team will finally build an autonomous facilitator leveraging a reinforcement learning model that optimizes for increasing group cohesion. The autonomous facilitator will be evaluated against a second agent that optimizes for equal access to the conversational floor, in terms of individual engagement and group cohesion assessed by post-session questionnaires. This work will build technologies for automated group facilitation that can assist to bridge the gaps in delivering support groups when human facilitators are absent or in short supply.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.