Conversational agents (CAs) have permeated our everyday lives in the past decade. Yet, the CAs we encounter today are far from perfect as they are still prone to breakdowns. Studies have shown that breakdowns have an immense impact on the user-CA relationship, user satisfaction, and retention. Therefore, it is important to investigate how to react and recover from breakdowns appropriately so that failures do not impair the CA experience lastingly. Examples for recovery strategies are the assumption of the most likely user intent (CA self-repair) or to ask for clarification (user-repair). In this paper, we iteratively develop a taxonomy to classify breakdown recovery strategies based on studies from scholarly literature and experiements with productive CA instances, and identify the current best practices described using our taxonomy. We aim to synthesize, structure and further the knowledge on breakdown handling and to provide a common language to describe recovery strategies.

Paper Number



Track 9: Human Computer Interaction & Social Online Behavior