Conversational Interfaces for Explainable AI: A Human-Centred Approach

Abstract

One major goal of Explainable Artificial Intelligence (XAI) in order to enhance trust in technology is to enable the user to enquire information and explanation directly from an intelligent agent. We propose Conversational Interfaces (CIs) to be the perfect setting, since they are intuitive for humans and computationally processible. While there are many approaches addressing technical and agent related issues of this human-agent communication problem, the user perspective appears to be widely neglected. With the goal of better requirement understanding and identification of implicit user expectations, a Wizard of Oz (WoZ) experiment was conducted, where participants tried to elicit basic information from a pretended artificial agent via Conversational Interface (What are your capabilities?). Chats were analysed by means of Conversation Analysis, where the hypothesis that users pursue fundamentally different strategies could be verified. Stated results illustrate the vast variety in human communication and disclose both requirements of users and obstacles in the implementation of protocols for interacting agents. Finally, we inferred essential indications for the implementation of such a CI. The findings show that existing intent-based design of Conversational Interfaces is very limited, even in a well-defined task-based interaction.