Design and Evaluation of a Conversational Agent to Support Domestic Violence Survivors

Domestic violence (DV) is widely recognized as a significant problem with detrimental impacts on the mental, physical, and socio-economic well-being of individuals, families, and the broader community. Despite various resources designated to support survivors, they may not be easily accessible or readily available. More importantly, DV survivors are often reluctant to divulge their experiences with others and may even refrain from seeking assistance because of social, emotional, privacy, and cultural concerns. As the immediate response to DV is crucial to survivors' physical and psychological well-being, they need prompt and non-blaming first responders.
Technological advancement, particularly in automated Conversational Agents (CAs), is progressing rapidly. CAs are gaining attention as a promising tool for providing counseling and support, addressing the above-mentioned challenges with survivors using traditional supporting resources. The primary objective of this dissertation research is to design, develop, and evaluate a CA-based solution that assists DV survivors in receiving support, enhancing their awareness, and increasing their access to services. To this end, we first identify the meta requirements and design principles of CA for survivors by interview-ing DV professionals, then design and develop SafeHaven, a CA-based prototype for supporting DV survivors, by following the design principles and meeting the meta requirements, and finally evaluate the effectiveness and user perception of SafeHavan by conducting user experiments.
Our findings suggest that CAs should empathize with survivors' experiences and provide them with meaningful informational, tangible, and emotional support, ensuring their safety as well as maintaining a transparent, private, and trustworthy dialogue. An evaluation of the CA was carried out with 36 participants, including DV survivors, their friends, family, and professionals, evaluating, among other metrics, the emotional, informational, and instrumental support provided. It was discovered that the CA outperformed traditional online searches and ChatGPT in terms of providing emotional, informational, and instrumental support, high information quality, and user trust.
This research identifies meta-requirements and design principles for designing a CA for DV survivors. This is the first research to evaluate the effectiveness of CAs in assisting individuals with DV, providing tailored, context-sensitive assistance that is superior to the capabilities of traditional search engines and general AI platforms like ChatGPT. In a broader sense, our results will be instrumental in guiding the development of future CA-driven support systems for DV survivors. This dissertation emphasizes the transformative potential of CAs for survivors of DV, as well as significant implications for CA developers, DV organizations, and support groups, proposing innovative strategies for enhancing anonymity, accessibility, and support for survivors.