Causal inference on strategies to increase conversational engagement for socially isolated older adults
Fengpei Yuan, Wenjun Zhou, Hiroko H Dodge, Tyler Morris, Xiaopeng Zhao- Psychiatry and Mental health
- Cellular and Molecular Neuroscience
- Geriatrics and Gerontology
- Neurology (clinical)
- Developmental Neuroscience
- Health Policy
- Epidemiology
Abstract
Background
Social isolation, characterized by a small social network and lack of contact, negatively impact the quality of life. Older adults are at an increased risk for social isolation, which is found to be a risk factor of cognitive decline and dementia. Enhancing social connectedness or interactions has been shown to be an effective intervention for slowing age‐associated cognitive decline and possibly delaying the onset of dementia. Here we aim to develop conversational chatbots which can provide engaging conversations with older adults with mild cognitive impairment (OAwMCI) to mitigate social isolation. To resemble human human interactions, chatbots have to adapt certain human strategies, such as empathy, during conversation.
Method
Our study utilized causal structural learning and causal inference approaches to examine the dialogue strategy for engaging OAwMCI. We analyzed transcriptions from a clinical trial, I‐CONECT (ClinicalTrials.gov #: NCT02871921), which involved semi‐structured conversations between facilitators and participants. By using emotion recognition in conversation and dialogue act modeling, we produced a time series data of state (participant’s emotions and length of utterance per conversation turn), intervention (facilitator’s conversational strategies), and outcome (participant’s following emotions and engagement). A R package, pcalg, was used to identify causal relationships between the state, intervention, and outcome variables.
Result
Graphical models implying causal relationships between certain factors were generated. Our analysis revealed that facilitators adjusted their dialogue strategies according to the participants' emotions and utterance. For example, facilitators expressed appreciation when a participant showed . Additionally, we found that the dialogue acts from facilitators influenced participants' emotions and engagement in the conversation. For instance, non‐opinion statements and yes/no questions from facilitators may affect participants’ verbal engagement and level of anger during the conversation.
Conclusion
Findings from this study are in line with real‐world human‐human conversation, and our next step is to explore how to use these findings to improve conversational chatbots. Future work shall also consider other crucial features into the state variable that might influence the dynamics of participants' status during the conversation, such as cognitive capacity and dialogue coherence. Our findings have the potential to inform the development of conversational tools and strategies to reduce social isolation among OAwMCI.