Plug and Play Conversations: The Micro-Conversation Scheme for Modular Development of Hybrid Conversational Agent

Conversational agents (CAs) for psychotherapy pose unique challenges (e.g., reliance on domain experts to pre-script large amounts of therapeutical dialogues). To tackle these challenges, we propose a modular approach to develop such CA, called Micro-Conversation Scheme (MCS). Conversations can be algorithmically extended in MCS by combining different micro-conversations (MC), which isolate single therapeutical topic. The sequencing of MC is managed by a Connector component, connecting MC into longer conversations with context. Additionally, MCS integrates natural language generation (NLG) models as plugins for generating counseling-style utterances (e.g., reflections). Moreover, MCS adopts interactive learning to continuously improve CA based on human feedback. MCS provides a solution to the challenges of complex-to-design and difficult-to-extend conversations, and inability of CA to flexibly generate context-appropriate counseling-style utterances for psychotherapy. MCS is expected to benefit the community by promoting the collaboration between conversational designers and developers while preserve their independence during the development of CAs.