Using meta-conversations and user data to improve the naturalness and user experience of DialoGPT
Summary
Negative feelings such as stress and anxiety are common in contemporary society, but not everybody can afford to get help from a professional. Several mental health coaching chatbots have appeared in the last few years, but they lack naturalness in conversation and understanding of context. On the other hand, general-purpose chatbots are not prepared to help users to navigate their negative feelings. Adding scripted conversations about feelings to general-purpose chatbots can create support chatbots for users looking to have a small talk with some friend to vent about their worries. Adding a meta-conversation (talking about the related conversation) about the emotions dialogue will give the users a deeper understanding of the workings of the chatbot and the usage of their data to personalise their experience. Meta-conversations are a great way to deepen the knowledge of a person about the discussed topic and the intention of their conversational partner, improving their relationship.
Prolonged exposure to the chatbot is necessary in order to gather enough data to personalise the user experience to test this chatbot. Therefore, 15 people participated in an experiment that lasted two weeks. After those two weeks, the participants filled out a questionnaire, and six of them participated in a subsequent interview. The results showed that the chatbot being incoherent complicates the interaction between the chatbot and the users.