Towards a comprehensive repair framework for human-chatbot interaction: the case of rephrasing

Chatbots are becoming a regular part of service offerings. However, failures in human-chatbot interactions are common and mitigating them with appropriate strategies is an integral part of the dialogue. For instance, chatbots can be designed to prompt a rephrase, although, due to the complexity of user reactions, this is not always successful. Research has called for taxonomies to categorize user reactions to compute meaningful responses that encourage dialogue continuation. We suggest a framework of strategies based on prior research and test its validity, focusing on how users rephrase across dialogue turns. We find that users rephrase problems formally (49%), by changing the number of words, altering syntax, or using synonyms and to a lesser extent by altering informational value (25%). We suggest training chatbots along this behavior and designing better prompts that guide users' next actions.