An Abduction-based Method for Explaining Non-entailments of Semantic Matching

Tracking #: 3837-5051

Authors: 
Ivan Gocev
Georgios Meditskos
Nick Bassiliades

Submission type: 
Full Paper
Abstract: 
Explainable Artificial Intelligence (XAI) attempts to give explanations for decisions made by AI systems. Despite the fact that for knowledge-based systems this is perceived as inherently easier than for black-box AI systems based on Machine Learning, research is still required for computing satisfactory explanations of reasoning results. In this paper, we focus on explaining non-entailments as a result of semantic matching in EL⊥ ontologies. In the cases where the result of semantic matching is an entailment, the already established methods of justifications and proofs provide excellent results. On the other hand, the cases in which the result of semantic matching presents in the form of a non-entailment demand an alternative approach. Inspired by abductive reasoning techniques, we present a method for computing subtree isomorphisms between graphical representations of EL⊥ concept descriptions, which are then used to construct solutions to abduction problems, i.e. explanations, for semantic matching non-entailments in EL⊥ ontologies. We improve existing results by generalizing our approach to be able to abduct complex concept expressions of all formats that also consist of role restrictions, rather than concepts alone, as well as the time needed to compute solutions to abduction problems in EL⊥ ontologies. We then illustrate our method with an example scenario and perform synthetic experiments to stress the methods’ capabilities and experiments on realistic ontologies to show the practical performance of our method.
Full PDF Version: 
Tags: 
Reviewed

Decision/Status: 
Accept

Solicited Reviews:
Click to Expand/Collapse
Review #1
Anonymous submitted on 09/May/2025
Suggestion:
Accept
Review Comment:

The overall quality of the paper has greatly improved and the comments have been addressed. Both the theoretical part and the experiments are more convincing, and I recommend acceptance of the paper.
I have only a few small comments:
- p.6, l.36: definition of \Pi^0: remove quantification over rho from definition (change to "...|v\in V, v!=\rho ..."), as \rho is defined to be the root node
- p.12,Def.11: the third and last bullet point seem to state the same thing and "\pi\in T" would be more like "\pi\in\Pi^0_T"
- p.13,l.5: "same root as T, thus \rho_S..."
- p.13: Proof of Observation 1 could be better structured
- p.14,l.40: "}"
- p.28,l.49: "...certain number of solutions that an abduction problem has", change wording
- p.29,l.43: "that have edges only labeled with one role", confusing wording
- p.32,l.1-5: Why does it occur mostly at around 1? Should it not increase gradually for smaller ratios?
- p.33,l.27-40: the third type of experiments need a more detailed explanation. E.g., why are there still so many trivial solutions?
- p.34,Tab.5: | instead of /

Review #2
Anonymous submitted on 26/May/2025
Suggestion:
Accept
Review Comment:

I am satisfied with this version of the submission, and I believe that it can be accepted.