Empirically Testing Explanation Preferences in Computational Argumentation

Computational argumentation is an AI approach for creating human-understandable systems and explanations. Within the computational argumentation literature, explanation definitions based on principles from cognitive science have been created to provide selective explanations, that is, to reduce the information included in explanations. However, it remains largely untested whether these explanations align with realistic explanation behaviour. In this study, we test and validate different explanation definitions based on computational argumentation. This is done through an experiment in which participants provide explanations for arguments by choosing from other presented arguments. Responses by participants are compared to three types of explanations based on the literature. This comparison leads us to conclude that people prefer short explanations consisting of related arguments. While these preferences partly align with argumentation-based explanation definitions, we also see that explanations are frequently shorter than those generated using these definitions. The results show that definitions from computational argumentation are indeed capable of providing human-like explanations, but that they could be further improved by creating explanation definitions that are even more selective.

The online version contains supplementary material available at https://doi.org/10.1007/s12559-026-10654-y.