Back to search

University of Minnesota

What to Say, What to Know: Enhancing User Modeling and Engagement with Conversational Agents

Abstract

dc:description.abstract

Users of conversational agents (CAs) frequently discontinue or severely limit their interactions with CAs over time. One of the biggest contributing factors to the decline in user acceptability is the user expectation versus reality gap. This gulf is thought to stem partly from anthropomorphization, which leads users to assume systems are much more socially capable than they actually are. This in turn leads to decreased acceptability when expectations of "humanness" go unmet. We propose that self-disclosure, or the act of revealing the self to others, is a key social phenomenon that may shape both initial user expectations and subsequent disappointment and declining use if CAs repeatedly fail to aptly respond to such revelations. This thesis uses a two-pronged approach to better understand self-disclosure as both a computational problem and a design tool. First, we explore self-disclosure as a challenge in natural language understanding. In preliminary experiments, we found that users frequently self-disclosed to a CA despite it being task-oriented and minimally social. Participants in other early work reported wanting personalized systems -- systems that can understand what a user discloses. We unified disparate literature to define a six-dimension ontology of self-disclosure and adapted the ontology to obtain human-rater evaluations of real user-CA dialogues containing user self-disclosure. These data revealed the dimension of inference, or the interpretation of disclosure, to be a particularly rich source of information, which we used to explore the performance of large language models (LLMs) as potential generators of inference data. Beyond language, the second prong of this thesis explores disclosure as an element of interaction design. We identified positive disclosure as a potential strategy to help boost user acceptance over time. To investigate its effectiveness, we designed and implemented a conversational strategy that deliberately encouraged users to participate in positive reflection and introspection. We created a prototype CA that utilized this strategy, and tested it with older adults over a week-long series of interactions. We found that the deliberate positivity probes were associated with sustained positive valence in user affect across time as well as a high degree of willingness to continue using the CA. These findings thus suggest that self-disclosure is not just a phenomenon to be expected –- it may also be an important part of a designer’s toolkit for both interaction and conversational design.

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Ferland, Elizabeth Alice Knouse

Subjects

dc:subject × 6

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Citation
Ferland, Elizabeth Alice Knouse. (2026). What to Say, What to Know: Enhancing User Modeling and Engagement with Conversational Agents. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/283083.

Chain of custody

source
Harvested from
University of Minnesota
Base URL
conservancy.umn.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Ferland, Elizabeth Alice Knouse. What to Say, What to Know: Enhancing User Modeling and Engagement with Conversational Agents. 2026. https://hdl.handle.net/11299/283083