Persona-based target-guided dialogue systems: A use case in dementia care

Abstract

In recent years, conversational agents have shown potential in various applications. However, the development of conversational agents tailored for older adults, particularly those with age-related cognitive limitations, remains unexplored. Inspired by person-centred care, this thesis proposes a framework for building a persona-based target-guided conversational agent. First, we train a BART model to computationally extract elements of a user’s ’persona’, which allows the agent to learn the individual’s background, preferences, and life stories. Then, we train a GPT-2 response generation model which leverages the extracted personas to generate personalised responses that preserve the identity of the individual, and train a keyword prediction kernel model to guide a conversation towards a given target topic. We evaluate the persona extraction model and response generation model on the public datasets PGDataset and ConvAI2 respectively. Then, we simulate conversations between a user agent and target-guiding agent to study the effects of incorporating persona information and keyword prediction into the conversational agent. Finally, we experiment the models on a new dataset constructed from a life-story interview transcript of older adults, and propose an experimental plan to evaluate the framework for people with dementia. Our long term goal is to address the challenges faced by individuals with dementia and their caregivers by building proactive person-centred conversational agents.