Development of a Human-Centred Psychometric Test for the Evaluation of Explanations Produced by XAI Methods

One goal of Explainable Artificial Intelligence (XAI) is to interpret and explain the inferential process of data-driven machine-learned models to make it comprehensible for humans. To reach it, it is necessary to have a reliable tool to collect the opinions of human users about the explanations generated by XAI methods of trained complex models. Psychometrics can be defined as the science behind psychological assessment. It studies the theory and techniques for measuring latent constructs such as intelligence, introversion, and conscientiousness. The knowledge developed in psychometrics was exploited to develop and evaluate a novel questionnaire for reliably evaluating the explanations produced by XAI methods. Explainability is a multi-faceted concept. Thus, it was necessary to create a set of questions to assess various facets and return a comprehensive, reliable measurement of explainability. The questionnaire development process was divided into two phases. First, a pilot study was designed and carried out to test the first version of the questionnaire. The results of this study were exploited to create a second, refined version of the questionnaire. The questionnaire was evaluated by assessing 1) its internal structure with the Exploratory Factor Analysis to analyse the interrelationships between the questionnaire’s items, 2) its reliability with the Cronbach alpha tests, and 3) its construct validity by comparing the distribution of the questionnaire’s answers with a set of quantitative metrics. Results showed that the questionnaire is promising as it was deemed a valid and reliable tool for evaluating XAI methods.