Erkan Basar, Divyaa Balaji and Linwei He
Conversational agents (chatbots) are increasingly applied on many areas, including personal health promotion and behaviour change. Health-related domains and behaviour change counselling benefit from using chatbot applications for their low-cost, 24/7 accessibility, and infinite patience towards its users. Such chatbots need to be able to engage users in a long-term interaction to achieve sustained healthy behaviour change. While doing so, they should be fair and non-judgmental against their users.
Methodologically, the current health-related chatbots often tend to be based on pre-scripted systems with rule-based and retrieval-based approaches. While these methods create highly controllable chatbots, their dialogues are often perceived as repetitive and mechanical. This results in reduced engagement with the chatbot. Conversely, the recent open-domain large language models demonstrated a potential for generating natural-sounding and coherent text. But they come with the cost of mishaps in communication and usage of inflammatory language, which renders end-to-end solutions too risky for an application in the health domain.
We introduce the LWT dialogue management system: a hybrid system attempting to combine the best parts of rule-based, retrieval-based, and generation-based approaches. The system uses a dialogue state tracker at its core to enforce a semi-predetermined dialogue flow. Each dialogue state can consist of multiple human-authored questions and responses, which can be tied to specific conditions. A retrieval model selects the top fitting responses from a set of candidates that pass their conditions. We use three generational models that generate various response candidates based on the conversational context. The set of generation-based candidates is combined with the set of human-authored candidates and given to the final response selection model, which decides on the utterance to be shared with the user.
In the Look Who’s Talking (LWT) project, we address the gap between the recent advancements in the computational linguistics area and real world applications by combining insights from artificial intelligence, health sciences, psychology and communication science. We aim to develop and test engaging chatbots for long-term serious conversations on health promotion, specifically in the two different behavioural change areas; smoking cessation and safe sex promotion.