Date: 27 March, 2024 @ 03:00PM- 04:00PM
Location: Online only (via Zoom)
Title: Conversational AI for Psychology and Mental Health
Speakers: Dan Lametti & Joanna Kuc
Abstract:
Conversational artificial intelligence is increasingly being used as a tool to engage research participants in and outside of the lab. In this talk, Dan Lametti, Associate Professor of Psychology at Acadia, will briefly review several studies run in collaboration with the artificial intelligence company OneReach.ai. In each case, conversational AI is used to address a different question in cognitive science and psychology. Joanna Kuc, PhD candidate in Experimental Psychology at University College London, will then present her work with OneReach examining how a Telegram-based conversational bot can be used to explore the relationship between language and mental health (abstract below). 100 English-speaking adults were recruited for a two-week exercise in which they journaled via a Telegram-based bot. Participants were divided into three groups. Each group received personalised feedback on their journal entries twice a week—either emotional content analysis or cognitive pattern recognition from the large language model GPT-4, or neutral acknowledgments. Across all groups, significant improvements in mental wellbeing were observed, with increases in wellbeing (WEMWBS) scores (+5.98, p<0.001), decreases in depression (PHQ-9) scores (-2.95, p<0.001), and reductions in anxiety (GAD-7) scores (-2.71, p<0.001). Engagement levels varied with feedback type. Participants who received an emotional analysis of their entries wrote more (avg. 2213 words) compared to a cognitive analysis (avg. 2015 words) or no analysis (avg. 1572 words). Linguistic content was analysed using RoBERTa- GoEmotions, a language model trained to classify 28 different emotions with probability scores. It was found that higher scores of depression and anxiety correlated with increased feelings of annoyance, sadness, disappointment, nervousness, and fear. Interestingly, depression also showed a stronger link with feelings of grief, whereas anxiety did not. Individuals with higher well being scores experienced less disapproval, annoyance, disgust, disappointment, and confusion, and notably, more joy. These findings offer insights into the interplay between personal language, emotional states, and mental health symptomatology. They also highlight the potential for integrating conversational AI into mental health monitoring.
Bio:
Dan Lametti is an Associate Professor of Cognitive Psychology at Acadia University and a Senior Advisor with OneReach.ai. Dan founded and runs the OneReach Academic Fellowship for conversational AI.
Joanna Kuc is a PhD student at University College London and OneReach.ai Academic Fellow. Her research focuses on naturalistic experience sampling through digital apps and wearables. Joanna is particularly interested in identifying linguistic and vocal biomarkers of mental health, and exploring how these are altered by psychedelic experiences.