Collaborative Research: Animated Agents in Self-Administered Surveys

本项目考察了动画或会话代理技术对基于网络的调查的影响。动画代理是用户界面中的拟人化软件对象，在其最先进的实现中，可以产生手势，显示面部表情，并通过语言协调移动眼睛。它们已经被证明可以提高用户在某些任务领域的表现（例如辅导）。这项技术什么时候会帮助或损害调查数据的质量和受访者的满意度？为了提供利益或伤害，代理人必须有多复杂？在网络调查中，代理可能会激励受访者参与并完成任务，就像人类采访者所做的那样，并帮助受访者理解调查问题，同时允许受访者在方便时参与，就像在普通的网络调查中一样。另一方面，代理的存在可能会阻碍对敏感话题的诚实回答，就像人类采访者所做的那样。在这个项目中，一系列的实验室实验检验了越来越多的类似人类的代理，这些代理会询问受访者的敏感和非敏感行为。这些研究将非代理网络调查中的数据质量和用户满意度与具有界面代理的数据质量和用户满意度进行了对比，这些界面代理的对话能力、提供有关其内部状态的视觉和口头线索的程度以及其讲话的意向性程度各不相同。代理人是用软件模拟的，该软件将现场面试官的视频图像实时转换为动画；因此，受访者认为他们正在与计算机生成的代理进行互动，尽管“代理”背后实际上有一个人。在关于非敏感行为的实验中，被调查者根据虚构的场景进行回答，以确定其回答的准确性。一个可能的结果是，具有更强对话能力的智能体将促进互动，从而导致更准确的理解，从而获得更准确的答案。在敏感行为实验中，被调查者回答自己的生活；敏感行为的报告越多，表明被访者越坦诚。一个可能的结果是，有更多动作的代理（嘴唇、眼睛和眉毛）会让受访者感到不那么隐私，因此比那些动作有限的代理回答得更不坦率。受访者的满意度是通过访谈后问卷来衡量的；不同的代理特征对受访者如何与调查系统沟通的影响是通过所有对话的详细逐轮编码来衡量的。这项工作的实际影响将是调查研究人员在采用代理技术时做出更明智的决策。了解代理何时提供帮助以及哪些特性帮助最大，可以集中精力决定开发哪些接口，不开发哪些接口。例如，如果对话能力对数据质量和用户满意度比其他代理特征更重要，那么这可能会将未来的开发工作重点放在代理的对话能力上，而不是视觉真实感上。拟议工作的理论影响将在两个方面。首先，它将加深我们对语言和非语言交流如何相互联系的理解，例如，语言互动如何受到代理人面部显示的保真度的影响。其次，通过比较人机交互和人机交互，该项目将推进关于意向性和人类代理的归因如何更普遍地影响交互的知识。作为支持调查和统计方法研究的联合活动的一部分，本研究得到了方法、测量和统计计划和联邦统计机构联盟的支持。

This project examines the impact of animated or conversational agent technology on web-based surveys. Animated agents are anthropomorphic software objects in the user interface that, in their most advanced implementations, produce gestures, display facial expressions, and move their eyes in coordination with their speech. They have been shown to improve user performance in some task domains (e.g. tutoring). When might this technology help or hurt survey data quality and respondents' satisfaction? How sophisticated must the agents be in order to provide benefit--or harm? In a web survey, an agent might motivate respondents to participate and complete the task, much like human interviewers do, and help respondents understand the survey questions as intended while allowing respondents to participate at their convenience, as in ordinary web surveys. On the other hand, the presence of an agent might discourage honest responding to questions about sensitive topics, much as human interviewers have been shown to do. In this project a series of laboratory experiments examine more and less human-like agents that ask questions about respondents' sensitive and non-sensitive behaviors. The studies contrast data quality and user satisfaction in non-agent web surveys to those with interface agents that vary in their dialogue capability, the degree to which they provide visual and spoken cues about their internal states, and the degree of intentionality in their speech. The agents are simulated with software that converts a video image of a live interviewer into an animation in real time; respondents thus believe they are interacting with a computer-generated agent even though there is actually a human behind the "agent." In the experiments about non-sensitive behaviors, respondents answer on the basis of fictional scenarios so that the accuracy of their answers can be determined. One possible outcome is that agents with greater dialogue capability will promote interactions that lead to more accurate understanding and thus more accurate answers. In the experiments about sensitive behaviors, respondents answer about their own lives; more reports of sensitive behaviors indicate greater respondent candor. One possible outcome is that agents with more movement (lips, eyes, and eyebrows) will lead respondents to feel less private and therefore to answer less candidly than with agents whose movement is limited. Respondents' satisfaction is measured with a post-interview questionnaire; the impact of different agent features on how respondents communicate with the survey system is measured by detailed turn-by-turn coding of all dialogue.The practical impact of this work will be more informed decisions by survey researchers in adopting agent technology. Knowing when agents help and what features help the most can focus decisions about what interfaces to develop and which ones not to develop. For example, if dialogue capability is more important to data quality and user satisfaction than other agent features, this could focus future development efforts on conversational competence of agents more than on visual realism. The theoretical impact of the proposed work will be in two areas. First, it will deepen our understanding of how verbal and non-verbal communication are interconnected, for example how verbal interaction is affected by the fidelity of the agents' facial display. Second, by comparing human-computer and human-human interaction the project will advance knowledge of how attributions of intentionality and human agency affect interaction more generally. This research is supported by the Methodology, Measurement, and Statistics Program and a consortium of federal statistical agencies as part of a joint activity to support research on survey and statistical methodology.