Plain Language to Address Dimensionality in Feature-Contribution Explanations for End-Users

Abstract

Prior works in XAI have attempted to create explanations that provide insight into how models make decisions, but few have focused on evaluating the utility of those explanations through user demographics and feedback. Moreover, bridging the gap between novice users and experts via tailored explanations is key to making explanations accessible to wider user groups. This work utilizes text readability and simplification tools to implement a plain language explanation that can be readily understood and evaluates user understanding by addressing both self-reported and objective understanding responses from participants. We assess this approach by applying the explanations in a data scalability context, where more features result in more information to communicate, complicating the ability to provide an effective explanation. Our findings showed dimensionality of the explanations had no impact on participant responses. Participant education level and familiarity with AI/ML emerged as significant factors, highlighting the need for further exploration of plain language as a method to craft informative yet approachable explanations for differing user backgrounds.