# Personally speaking: Developing and evaluating an ontology of dimensions of meaning for self-disclosure with a conversational assistant

## TL;DR

This paper introduces an ontology to better understand and interpret self-disclosure in conversations with AI assistants, aiming to improve long-term user engagement.

## Contribution

The novel contribution is an ontology that integrates and expands dimensions of self-disclosure, including new layers like the separation of topic from mental verbs.

## Key findings

- The ontology captures dimensions like topic and intimacy, along with new layers such as habit vs. preference vs. memory.
- Application of the ontology to real dialogues reveals relationships between different dimensions of self-disclosure.
- The findings suggest practical implications for improving conversational agents through better modeling of self-disclosure.

## Abstract

Dialogue systems and conversational assistants are promising technologies given their general accessibility and appeal, but longer-term adoption often falters. Creating systems that engage users over the long term is a challenging design problem, largely because it depends on relationship formation between user and agent. Self-disclosure, or the act of revealing information about oneself, is a fundamental component of relationship building and maintenance between humans, and it has been shown to occur in interactions between humans and language-based systems as well. Disclosure on the part of users is an exceptionally rich source of information that has the potential to shape everything from user modeling to conversational experience design; however, that same richness makes interpreting disclosure difficult. Although some research has examined different sources of meaning such as topic and intimacy, the convergence of these sources of meaning under one umbrella has yet to be considered. We propose an ontology of self-disclosure with dialogue systems as a means to address this gap. The proposed ontology encapsulates previously explored dimensions of self-disclosure, such as topic and intimacy, as well as some additional novel layers of meaning, such as the separation of topic from the mental verb referred to in the disclosure (e.g., habit vs. preference vs. memory), in order to further discretize the separate dimensions of this complex phenomenon and make explicit potentially valuable sources of information for agents. We demonstrate an application of this ontology to instances of self-disclosure, drawn from real dialogues between users and a task-oriented conversational assistant, and examine the observed relationships between different dimensions of meaning. The practical implications of these findings, as well as the potential for further developing the ontology, demonstrate the usefulness and value of approaching self-disclosure as a multi-faceted, interconnected phenomenon.