PhD Dissertation Proposal Defense: Mashrur Rashik, Human-Centered Design of Contextually Appropriate Conversational Agents for Enhancing Public Interaction, Engagement, and Multimodal Experiences

Content

Speaker: Mashrur Rashik
Advisors: Narges Mahyar and Ali Sarvghad

Abstract: In recent years, the popularity of AI-enabled conversational agents or chatbots has risen as an alternative to traditional online surveys to elicit information from people. Despite their growing popularity, current conversational agents frequently struggle to understand user intentions during complex, multifaceted conversations, and they lack comprehensive design considerations—particularly in avatar representation—that influence user engagement and trust. In addition, traditional data collection methods often rely on static, self-directed entries, lacking interactive feedback and real-time guidance. This gap can lead to incomplete or imprecise data, limiting its usefulness in specialized domains like healthcare, where healthcare providers use the collected data to provide patient care. In this dissertation, I investigate the design, development, and evaluation of conversational agents to facilitate engaging user interactions and rich information elicitation.

First, I investigated how multi-agent chatbot systems can collect data from multifaceted conversations spanning multiple domains. To that end, I conducted a Wizard of Oz study to examine the design of a multi-agent chatbot for gathering public input across multiple high-level domains and their associated topics. Building on the findings, I designed and developed  CommunityBots—a multi-agent chatbot platform where each chatbot agent handles a different domain individually. To manage conversation across multiple topics and chatbots, I introduced a novel Conversation and Topic Management (CTM) mechanism that handles topic-switching and chatbot-switching based on user responses and intentions. A between-subjects study with 96  crowd workers demonstrated that CommunityBots significantly increased user engagement, improved response quality, and reduced conversational interruptions compared to a single-agent baseline. The visual cues integrated into the interface helped participants better understand the functionalities of the CTM mechanism, resulting in increased user satisfaction.

Next, to gain a holistic understanding of conversational agent avatar design space, I conducted a comprehensive analysis of existing literature to map avatar design space. I defined a categorization of 10 dimensions that is based on the analysis and iterative coding of 266 conversational agent papers from 160 venues spanning 2003 to the present. In addition, I built an interactive browser to facilitate exploration and interaction with these dimensions and their interrelationships. This categorization lays the  
groundwork for researchers, designers, and practitioners to discern task-specific and contextual aspects of conversational agent avatar design.

Finally, I explored the application of conversational agents in healthcare by introducing PATRIKA, an AI-enabled conversational journaling prototype designed for people with Parkinson's disease (PwPD). Traditional journaling methods, such as online surveys, often lack interactive feedback and real-time guidance, leading to incomplete or imprecise data, particularly for  
managing chronic conditions. To address this gap, PATRIKA incorporates cooperative conversation principles, clinical interview simulations, and personalization to create a more engaging and effective journaling experience. Through two user studies with PwPD and iterative refinements, I demonstrated that conversational journaling significantly enhances patient engagement and collects clinically valuable information. My analysis revealed that by generating probing questions, PATRIKA transformed journaling into a bi-directional interaction, providing useful insights for designing future journaling systems in healthcare.

I am currently working on personalized conversational agents to improve engagement and learning about climate change. Traditional climate change learning methods often rely on static visualizations that may not accommodate diverse learning needs. To address this challenge, I am working on a personalizable conversational agent that tailors its conversation style, personality, and visualization choices to individual users. The agent dynamically adjusts explanations based on users' data visualization  
background and geographical location. This dissertation contributes to our understanding of conversational agent design by addressing key issues in multi-agent systems, avatar design, and healthcare journaling. The findings contribute to the development of more user-centered conversational agents, with implications for public input elicitation, user engagement, and rich data collection.

Join the Zoom