Seeing us through machines: designing and building conversational AI to understand humans

Xiao, Ziang

Permalink

https://hdl.handle.net/2142/120280 Copy

Description

Title

Seeing us through machines: designing and building conversational AI to understand humans

Author(s)

Xiao, Ziang

Issue Date

2023-04-21

Director of Research (if dissertation) or Advisor (if thesis)

• Sundaram, Hari
• Karahalios, Karrie

Doctoral Committee Chair(s)

• Sundaram, Hari
• Karahalios, Karrie

Committee Member(s)

• Zhou, Michelle X.
• Ji, Heng
• Roberts, Brent W.

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Understanding Human
• Conversational Ai
• Artificial Intelligence
• Human-computer Interaction
• Natural Language Processing
• Survey
• Informed Consent
• Voice Interaction
• Personality

Language

eng

Abstract

Understanding humans at scale is essential for addressing some of the consequential challenges in human society. By gaining insights into why people act as they do, we can design informed interventions that have a positive societal impact, including improving public health, developing a sustainable economy, or advancing fair education. However, the complexity of human behavior necessitates novel and sophisticated tools and methods to capture cultural, social, environmental, and individual characteristics that heavily influence our behaviors. Further, understanding ethical concerns surrounding informed consent, privacy, and data collection requires interdisciplinary expertise and is essential to study human behavior responsibly. In this dissertation, we take up this challenge by exploring the use of Artificial Intelligence (AI) in the context of behavioral science studies, designing and building effective conversational AIs for information collection and informed consent. This thesis starts by focusing on surveys, one of the most widely-used research methods in behavioral research. We studied conversational AIs to address today's survey research challenges: survey fatigue, inflexible survey structure, and lack of personalization. In an AI-driven conversational survey, a conversational agent asks questions, interprets a participant's responses, and probes answers whenever needed. We first studied an AI-driven conversational survey's response quality and participant engagement by comparing it with form-based surveys. After establishing the promise, We improved a conversational survey by equipping the AI agent with active listening skills through a human-in-the-loop framework. We further built a novel knowledge-driven language model to generate informative follow-up questions on the fly. We then looked at the ethical practices in behavioral science research, informed consent procedure, in online studies. Due to the lack of a researcher's presence and guidance, online participants often failed to make informed participation decisions, putting them at unaware risks. In this study, we re-introduced interactivity to online informed consent using conversational AI. Our agent guided participants through the consent form step-by-step and answered their questions. Compared to the form-based interaction, we found the AI-powered chatbot improved consent form reading, promoted participants’ feelings of agency, closed the power gap between the participant and the researcher, and ultimately benefited the study quality. We ended this thesis with a series of empirical studies about how people interact with such conversational AIs. Drawing from the rich use of voice assistants, we considered voice as another modality. We studied how voice assistants with different social metaphors influence people's reactions and perceptions of their information requests. We then deployed a conversational AI in the real world to collect students' team preferences and demonstrated how such an agent improves the student teaming experience. This dissertation provides both empirical evidence of how to design effective conversational AI to understand human behavior at scale and technical frameworks to build such an agent. Most importantly, it contributes to design implications for future technologies to improve our understanding of how we interact with each other and our environment and push this research field forward.

Graduation Semester

2023-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/120280

Copyright and License Information

Copyright 2023 Ziang Xiao

Owning Collections