Conversational XAI: Formalizing Its Basic Design Principles

eXplainable Artificial Intelligence (XAI) aims to explain the predictions and operations performed by an AI model. Its goal is to make AI models more understandable to humans. However, XAI methods sometimes produce explanations in implementation-dependent formats and these artifacts may stimulate different perceptions in users with different backgrounds. Conversational XAI systems have been proposed to provide explanations in the form of conversation based on natural language. This new trend for XAI systems focused on a human-centered approach provides more powerful forms of explanation representation. In this study, we analyze the current state of the art of Conversational XAI systems and propose a general formalization based on currently available literature. Moreover, we devise a general Conversational XAI architecture that includes two new components designed to improve the user experience both functionally taking into account the recurrent questions and in terms of trustworthiness by explicitly providing metrics for the explanation.