The Dual Role of Abstracting over the Irrelevant in Symbolic Explanations: Cognitive Effort vs. Understanding

Explanations are central to human cognition, yet AI systems often produce outputs that are difficult to understand. While symbolic AI offers a transparent foundation for interpretability, raw logical traces often impose a high extraneous cognitive load. We investigate how formal abstractions, specifically removal and clustering, impact human reasoning performance and cognitive effort. Utilizing Answer Set Programming (ASP) as a formal framework, we define a notion of irrelevant details to be abstracted over to obtain simplified explanations. Our cognitive experiments, in which participants classified stimuli across domains with explanations derived from an answer set program, show that clustering details significantly improve participants’ understanding, while removal of details significantly reduce cognitive effort, supporting the hypothesis that abstraction enhances human-centered symbolic explanations.

💡 Research Summary

The paper tackles a fundamental challenge in explainable artificial intelligence: while symbolic AI promises transparency, the raw logical traces it produces often overwhelm human users with excessive detail, leading to high cognitive load. To address this, the authors investigate two concrete abstraction techniques—removal and clustering—and assess how each influences human understanding and effort when explanations are derived from Answer Set Programming (ASP) models.

First, the authors formalize a notion of “χ‑irrelevance” that generalizes earlier, more restrictive abstraction concepts. Given a program P and a set of potential problem instances χ, χ‑irrelevance requires the existence of a surjective mapping m from the original atom universe to an abstracted universe (or to logical truth ⊤) such that, for every instance F in χ, the abstracted answer sets of P∪F coincide with the answer sets of the mapped program m(P)∪m(F). The mapping m embodies two elementary cognitive operations: (i) removal, which maps irrelevant atoms to ⊤, effectively forgetting details that do not affect the decision; and (ii) clustering, which maps several distinct atoms to a single representative atom, thereby reducing granularity and “chunking” related concepts. This definition guarantees that the abstracted program preserves the solution space across all considered scenarios, making it suitable for generating concise justifications.

To illustrate the approach, the authors use the xclingo tool to visualize explanations for a toy flower‑classification task. In that example, the attribute “spiky” is true for all flowers and thus removable, while the habitats “water” and “mud” both lead to the same downstream predicate “needsWater” and can be clustered into a single abstract concept. The resulting abstracted program is dramatically shorter yet semantically equivalent for the instances under study.

The empirical component consists of a user study involving three biologically inspired domains—flowers, mushrooms, and cacti—each defined by six domain‑specific attributes and a binary target label. Participants were presented with explanations of four types: (1) the original, unabstracted trace; (2) removal‑only; (3) clustering‑only; and (4) a combination of removal and clustering. After viewing an explanation, participants performed a binary classification on novel instances. The study measured (a) accuracy (as a proxy for understanding), (b) response time (as a proxy for cognitive effort), and (c) self‑reported confidence.

Results reveal a double‑edged benefit of abstraction. Clustering significantly improves accuracy, indicating that grouping related details helps participants grasp the essential logical structure. Removal, on the other hand, markedly reduces response times, confirming that eliminating irrelevant facts eases the parsing burden. The combined abstraction (both removal and clustering) yields the best overall performance: highest accuracy, fastest responses, and the greatest confidence boost. These findings support the hypothesis that well‑designed abstractions can simultaneously enhance comprehension and efficiency.

The authors conclude that abstraction is not an optional embellishment but a necessary component of human‑centric symbolic explanations. By formally defining and empirically validating removal and clustering within ASP, the work demonstrates that complex AI reasoning can be rendered both transparent and cognitively manageable. Future directions include adaptive abstraction levels, integration with neuro‑symbolic architectures, and broader domain evaluations, aiming to further bridge the gap between formal AI reasoning and human interpretability.