Abduction under Repair Semantics

• Abduction under Repair Semantics is a nonmonotonic explanatory framework that evaluates hypotheses against repairs of inconsistent knowledge bases rather than the raw data.
• It distinguishes between brave and AR semantics, with brave requiring validation in at least one repair and AR demanding consistency across all repairs.
• Research advances focus on minimality, conflict-confinement, and signature restrictions to ensure explanations remain meaningful and computationally tractable.

Abduction under repair semantics is a form of non-monotonic explanatory reasoning for inconsistent knowledge bases in which hypotheses are evaluated not against the raw inconsistent theory, but against its repairs. In the recent description-logic formulation, the starting point is an inconsistent knowledge base $\mathcal K=\langle \mathcal T,\mathcal A\rangle$ together with an observation $\alpha$ that is not entailed under a chosen repair semantics; a hypothesis is then an ABox extension $H$ such that $$\mathcal T,\mathcal A\cup H\models_{\mathcal S}\alpha$$ for $\mathcal S\in\{\mathrm{brave},\mathrm{ar}\}$ (Haak et al., 29 Jul 2025). This departs from classical abduction, where consistency of the augmented knowledge base is required explicitly, and it also differs from database-repair approaches that focus on restoring global consistency rather than explaining a missing entailment. The topic lies at the intersection of repair theory, inconsistency-tolerant query answering, and logic-based abduction, with early model-theoretic and abductive work on repairing inconsistent databases providing a precursor perspective [0207085].

1. Historical and conceptual origins

A foundational precursor is the model-theoretic analysis of inconsistent databases by Arenas, Bertossi, and Chomicki’s contemporaries in the early repair literature, and, in the specific abductive line considered here, the paper "Repairing Inconsistent Databases: A Model-Theoretic Approach and Abductive Reasoning" [0207085]. That work stated two complementary perspectives: a model-theoretic characterization of ways to "repair" a database by recovering consistent data from an inconsistent database, and an abductive application based on an abductive solver, the A-system, implementing an SLDNFA-resolution procedure, which computes data-facts to be inserted or retracted so as to keep the database consistent [0207085]. The abstract also states that the model-theoretic and abductive approaches are connected by soundness and completeness results.

In later description-logic work, the emphasis shifted from repairing inconsistency itself to explaining missing entailments in the presence of inconsistency. The paper "Why not? Developing ABox Abduction beyond Repairs" (Haak et al., 29 Jul 2025) explicitly positions the problem as abduction for inconsistent KBs under repair semantics, while "ABox Abduction for Inconsistent Knowledge Bases under Repair Semantics" (Haak et al., 2 May 2026) provides a broader complexity map for this setting. A related but conceptually distinct line views abduction as a mechanism for KB repair in the sense of supplying missing assertions to recover an intended entailment, not in the sense of defining a repair semantics for inconsistent KBs (Koopmann, 2021).

This suggests a useful distinction between two research threads. One thread studies repair as the primary semantic object, so that reasoning proceeds over maximal consistent fragments or related structures. The other studies abduction as an explanatory operation whose output may itself function as a repair of a missing entailment. The topic of abduction under repair semantics occupies the overlap: the semantics is repair-based, but the computational task is abductive.

2. Repair semantics as the reasoning substrate

The standard setting in the recent description-logic papers assumes a knowledge base $\mathcal K=\langle \mathcal T,\mathcal A\rangle$ with inconsistent ABox data. A repair is a subset $\mathcal R\subseteq\mathcal A$ such that $\langle\mathcal T,\mathcal R\rangle\not\models\bot$ and $\mathcal R$ is subset-maximal with this property; a conflict is a subset-minimal $\mathcal T$-inconsistent subset of the ABox (Haak et al., 29 Jul 2025). The set of all conflicts is denoted $\alpha$0 (Haak et al., 29 Jul 2025).

Two semantics dominate this literature. Under brave semantics,

$\alpha$1

Under AR semantics,

$\alpha$2

These definitions are treated as standard inconsistency-tolerant semantics in the recent ABox-abduction papers (Haak et al., 29 Jul 2025).

The repair layer restores non-trivial reasoning because classical entailment over an inconsistent KB becomes uninformative. "ABox Abduction for Inconsistent Knowledge Bases under Repair Semantics" (Haak et al., 2 May 2026) states this point directly: with inconsistency, classical entailment collapses, and repair semantics regains meaningful reasoning by reasoning over maximal consistent ABox fragments. In that paper, the baseline entailment complexity recalled for concept assertions is $\alpha$3-complete for $\alpha$4 entailment in $\alpha$5 and $\alpha$6-complete in $\alpha$7, while $\alpha$8-entailment is $\alpha$9-complete for all three logics considered there (Haak et al., 2 May 2026).

A related but different repair semantics appears in "Query Answering with Inconsistent Existential Rules under Stable Model Semantics" (Wan et al., 2016). There, the database is assumed reliable and the rules may be unreliable; a repair is therefore a preferred subset $H$0 of the rule set such that $H$1 has a stable model, and no strictly preferred superset does (Wan et al., 2016). Query answering requires truth in all preferred rule repairs. The paper explicitly connects these preferences to logic-based abduction, interpreting a rule repair as a preferred explanatory hypothesis about which rules to keep (Wan et al., 2016). This does not define ABox abduction under repair semantics in the newer DL sense, but it shows that repair semantics can shift from data repair to hypothesis selection over rules.

3. Formal definition of abduction under repair semantics

The recent DL formulation begins with a promise setting. Let $H$2 be inconsistent, let $H$3 be an atomic Boolean instance query or concept assertion, and let $H$4 such that $H$5; then $H$6 is an $H$7-abduction problem (Haak et al., 29 Jul 2025). A hypothesis is an ABox $H$8 such that

$H$9

In the 2026 formulation, hypotheses are required to use only individuals occurring in $$\mathcal T,\mathcal A\cup H\models_{\mathcal S}\alpha$$0 and $$\mathcal T,\mathcal A\cup H\models_{\mathcal S}\alpha$$1, and a signature-restricted variant $$\mathcal T,\mathcal A\cup H\models_{\mathcal S}\alpha$$2 further requires $$\mathcal T,\mathcal A\cup H\models_{\mathcal S}\alpha$$3 to use only symbols from $$\mathcal T,\mathcal A\cup H\models_{\mathcal S}\alpha$$4 (Haak et al., 2 May 2026).

The decisive difference from classical abduction is that the augmented KB need not be classically consistent. Classical ABox abduction asks for $$\mathcal T,\mathcal A\cup H\models_{\mathcal S}\alpha$$5 such that $$\mathcal T,\mathcal A\cup H\models_{\mathcal S}\alpha$$6 and $$\mathcal T,\mathcal A\cup H\models_{\mathcal S}\alpha$$7; under repair semantics, the full KB may remain inconsistent, and the target condition is instead repair-based entailment (Haak et al., 2 May 2026). In the consistent case, these notions collapse because there is essentially only one repair, namely the whole ABox; the distinctive phenomena arise only in the inconsistent setting (Haak et al., 29 Jul 2025).

A central structural phenomenon is the role of the trivial hypothesis $$\mathcal T,\mathcal A\cup H\models_{\mathcal S}\alpha$$8. Because observations are atomic BIQs, $$\mathcal T,\mathcal A\cup H\models_{\mathcal S}\alpha$$9 can itself be a candidate explanation. The papers state that under brave semantics, $\mathcal S\in\{\mathrm{brave},\mathrm{ar}\}$0 is always a hypothesis; under AR semantics, $\mathcal S\in\{\mathrm{brave},\mathrm{ar}\}$1 may or may not work (Haak et al., 29 Jul 2025). The 2026 paper formulates this as a lemma: for brave abduction, there is a brave-hypothesis for $\mathcal S\in\{\mathrm{brave},\mathrm{ar}\}$2 iff $\mathcal S\in\{\mathrm{brave},\mathrm{ar}\}$3 is a brave-hypothesis; for AR abduction, there is an AR-hypothesis iff $\mathcal S\in\{\mathrm{brave},\mathrm{ar}\}$4 is an AR-hypothesis iff $\mathcal S\in\{\mathrm{brave},\mathrm{ar}\}$5 is conflict-confining for $\mathcal S\in\{\mathrm{brave},\mathrm{ar}\}$6 (Haak et al., 2 May 2026). This is one of the main reasons why non-triviality and signature restrictions become important.

A common misconception is to identify this framework with ordinary repair of inconsistent data. The papers are explicit that the task is not to compute a repair of $\mathcal S\in\{\mathrm{brave},\mathrm{ar}\}$7 itself, but to compute a hypothesis that makes a missing entailment true under repair semantics. By contrast, the 2002 database paper uses abduction to compute insertions or retractions that restore consistency directly [0207085]. The two viewpoints are related, but they address different targets.

4. Hypothesis properties and repair-aware minimality

Because unrestricted repair-semantic abduction admits trivial or uninformative solutions, the literature develops several criteria for "useful" hypotheses. The 2026 paper lists four main properties: non-trivial, meaning $\mathcal S\in\{\mathrm{brave},\mathrm{ar}\}$8; $\mathcal S\in\{\mathrm{brave},\mathrm{ar}\}$9-minimal for $\mathcal K=\langle \mathcal T,\mathcal A\rangle$0; conflict-confining, meaning

$\mathcal K=\langle \mathcal T,\mathcal A\rangle$1

and $\mathcal K=\langle \mathcal T,\mathcal A\rangle$2-minimal, meaning that no hypothesis yields a strictly smaller conflict set than $\mathcal K=\langle \mathcal T,\mathcal A\rangle$3 (Haak et al., 2 May 2026).

Conflict-confinement is the most repair-specific notion. In the 2025 paper it is defined by

$\mathcal K=\langle \mathcal T,\mathcal A\rangle$4

and the paper notes that this is equivalent to

$\mathcal K=\langle \mathcal T,\mathcal A\rangle$5

Thus, adding the hypothesis must not create any new conflicts relative to the original inconsistent KB (Haak et al., 29 Jul 2025). The 2026 paper generalizes this to conflict-minimality, allowing some new conflicts but penalizing hypotheses that introduce more of them (Haak et al., 2 May 2026).

The papers also study signature-based restrictions. In the more general ABox-abduction literature, a problem is given as $\mathcal K=\langle \mathcal T,\mathcal A\rangle$6, and a hypothesis $\mathcal K=\langle \mathcal T,\mathcal A\rangle$7 must satisfy $\mathcal K=\langle \mathcal T,\mathcal A\rangle$8 together with consistency and entailment conditions (Koopmann, 2021). That work stresses that signature restriction preserves explanatory character by excluding "too shallow" or irrelevant names (Koopmann, 2021). In the repair-semantics setting, the same device serves an additional role: it blocks the trivial hypothesis $\mathcal K=\langle \mathcal T,\mathcal A\rangle$9 when the observation symbol is not abducible (Haak et al., 29 Jul 2025).

Another line of work develops stronger relevance criteria in ontology-repair settings. In TBox abduction for $\mathcal R\subseteq\mathcal A$0, "Connection-minimal Abduction in EL via Translation to FOL" (Haifani et al., 2022) argues that subset, size, and semantic minimality are insufficient because they may allow hypotheses using concepts unrelated to the problem at hand. It introduces connection minimality, defined through concepts $\mathcal R\subseteq\mathcal A$1 that connect the left- and right-hand sides of the observation and a weak homomorphism between their description trees (Haifani et al., 2022). Although this is TBox abduction rather than ABox abduction under repair semantics, it clarifies a general methodological point: minimality in repair-oriented abduction is not only about size, but also about structural relevance.

5. Complexity landscape

The most systematic classifications are given for $\mathcal R\subseteq\mathcal A$2 and $\mathcal R\subseteq\mathcal A$3 in (Haak et al., 29 Jul 2025, Haak et al., 2 May 2026), and (Haak et al., 17 Jun 2026). The results show that complexity depends sharply on the repair semantics, the logic, and the hypothesis property imposed.

Before summarizing the numbers, two structural explanations recur. First, in $\mathcal R\subseteq\mathcal A$4, conflicts have size at most $\mathcal R\subseteq\mathcal A$5, and minimal $\mathcal R\subseteq\mathcal A$6-supports of concept assertions have size $\mathcal R\subseteq\mathcal A$7 (Haak et al., 2 May 2026). Second, the set of AR-hypotheses need not be convex: there may be $\mathcal R\subseteq\mathcal A$8 with $\mathcal R\subseteq\mathcal A$9 hypotheses but $\langle\mathcal T,\mathcal R\rangle\not\models\bot$0 not (Haak et al., 2 May 2026). The former tends to keep complexity low in $\langle\mathcal T,\mathcal R\rangle\not\models\bot$1; the latter helps explain higher complexity under AR semantics, especially with minimality.

Setting	Main existence results	Main verification results
$\langle\mathcal T,\mathcal R\rangle\not\models\bot$2 under repair semantics	General hypotheses: $\langle\mathcal T,\mathcal R\rangle\not\models\bot$3-complete for $\langle\mathcal T,\mathcal R\rangle\not\models\bot$4 and $\langle\mathcal T,\mathcal R\rangle\not\models\bot$5; signature-restricted: $\langle\mathcal T,\mathcal R\rangle\not\models\bot$6-complete for $\langle\mathcal T,\mathcal R\rangle\not\models\bot$7 and $\langle\mathcal T,\mathcal R\rangle\not\models\bot$8-complete for $\langle\mathcal T,\mathcal R\rangle\not\models\bot$9 (Haak et al., 2 May 2026)	General and $\mathcal R$0-minimal: $\mathcal R$1-complete for $\mathcal R$2 and $\mathcal R$3-complete for $\mathcal R$4; $\mathcal R$5-minimal AR: $\mathcal R$6-complete (Haak et al., 2 May 2026)
$\mathcal R$7 under repair semantics	General hypotheses: $\mathcal R$8-complete for $\mathcal R$9 and $\mathcal T$0-complete for $\mathcal T$1; signature-restricted: $\mathcal T$2-complete for $\mathcal T$3 and $\mathcal T$4-complete for $\mathcal T$5 (Haak et al., 2 May 2026)	General and $\mathcal T$6-minimal: $\mathcal T$7-complete for $\mathcal T$8 and $\mathcal T$9-complete for $\alpha$00; $\alpha$01-minimal: $\alpha$02-complete for $\alpha$03 and $\alpha$04-complete for $\alpha$05 (Haak et al., 2 May 2026)
$\alpha$06 with combined properties	Conflict-confining non-trivial existence: $\alpha$07-complete for both $\alpha$08 and $\alpha$09 (Haak et al., 17 Jun 2026)	Conflict-confining non-trivial verification: $\alpha$10-complete for $\alpha$11 and $\alpha$12-complete for $\alpha$13; non-trivial or signature-restricted $\alpha$14-minimal verification: $\alpha$15-complete for both semantics (Haak et al., 17 Jun 2026)

The 2025 paper had already established core results in a slightly narrower setting. Without signature restriction, existence is trivial under brave semantics because $\alpha$16 is always a brave-hypothesis, whereas under AR semantics existence is equivalent to checking whether $\alpha$17 is conflict-confining, equivalently whether $\alpha$18; the paper gives $\alpha$19-completeness for $\alpha$20 and $\alpha$21-completeness for $\alpha$22 (Haak et al., 29 Jul 2025). With signature restriction, existence becomes $\alpha$23-complete for brave and $\alpha$24-complete for AR in $\alpha$25, while for $\alpha$26 the paper gives $\alpha$27-completeness for brave and $\alpha$28-membership for AR, with exact completeness left open there (Haak et al., 29 Jul 2025).

The 2026 follow-up on combining properties in $\alpha$29 emphasizes a meta-pattern: adding signature restriction or non-triviality often does not increase complexity beyond the corresponding harder base property, because these properties are polynomial-time checkable and downward closed (Haak et al., 17 Jun 2026). The paper’s “main observation” is that often requiring additional properties for hypotheses does not lead to an increase of complexity (Haak et al., 17 Jun 2026).

A plausible implication is that the computational bottlenecks in repair-semantic abduction are driven less by simple syntactic admissibility constraints and more by the interaction between repair quantification and optimality criteria such as subset-minimality or conflict-minimality.

6. Related formulations and broader significance

The topic connects to several neighboring areas that use abduction for repair-like purposes without adopting the same formal semantics. In signature-based ABox abduction for expressive description logics, hypotheses may require fresh individuals, complex concepts, or both; the paper "Signature-Based Abduction with Fresh Individuals and Complex Concepts for Description Logics" (Koopmann, 2021) studies the complexity and size of such explanations. It states that ABox abduction has applications such as diagnosis, KB repair, and explaining missing entailments, but also makes clear that it does not define repairs, minimal repairs, or inconsistency-repair semantics (Koopmann, 2021). Its repair connection is therefore explanatory rather than semantic.

That paper also shows how strongly syntactic restrictions affect the feasibility of explanation. For flat hypotheses, every solvable instance has a polynomial-size hypothesis in $\alpha$30, an exponential-size hypothesis in $\alpha$31 and $\alpha$32, and no general upper bound in $\alpha$33; with complex concepts, smallest hypotheses can be triple exponential in $\alpha$34 (Koopmann, 2021). This suggests that if repair-semantic abduction is extended beyond flat ABoxes and lightweight DLs, the choice of hypothesis language may become a decisive complexity parameter.

The rule-repair framework under stable model semantics offers another perspective on abductive repair. There, the hypothesis space is not a set of ABox assertions but a set of rules to retain. Preferred repairs are selected by inclusion, cardinality, priorities, or weights, and the paper explicitly states that these preference schemes are inspired by abduction and preferred explanation selection (Wan et al., 2016). This can be read as an abductive semantics of rule reliability rather than a semantics of ABox extension.

Across these variants, one recurring theme is that repair-based abduction is best understood as a family of methods for reasoning under defeasibility induced by inconsistency. In one variant, the defeasible objects are ABox assertions retained in repairs; in another, they are candidate hypotheses added to an already inconsistent KB; in another, they are rules whose reliability is under question (Wan et al., 2016). What remains constant is the shift from classical monotonic entailment to reasoning over preferred consistent surrogates.

7. Characteristic phenomena and research directions

Several phenomena distinguish abduction under repair semantics from classical abduction. First, the observation itself may be a hypothesis, especially under brave semantics (Haak et al., 29 Jul 2025). Second, hypotheses need not preserve consistency of the full augmented KB; what matters is entailment over repairs (Haak et al., 2 May 2026). Third, conflict-sensitive criteria such as conflict-confinement and $\alpha$35-minimality have no direct classical counterpart and are specific to inconsistent-data settings (Haak et al., 29 Jul 2025). Fourth, minimality behaves differently because AR-hypotheses can be non-convex, which contributes to higher complexity for subset-minimal verification (Haak et al., 2 May 2026).

The recent literature also indicates which questions remain open or only partially settled. In the 2025 study, the exact completeness of signature-restricted AR-existence in $\alpha$36 remained open, with $\alpha$37-membership shown (Haak et al., 29 Jul 2025). In the 2026 combined-properties paper, the exact $\alpha$38 complexity of verification for $\alpha$39-minimal conflict-confining hypotheses is left open, and the paper also notes that the cardinality-conflict-minimal case $\alpha$40 is potentially much harder because a small hypothesis may induce exponentially many conflicts (Haak et al., 17 Jun 2026).

A broader research direction concerns integration with more expressive explanation languages. The lightweight-logics results are comparatively well mapped, whereas the expressive-DL literature already shows that allowing fresh individuals and complex concepts can produce exponential, double exponential, or triple exponential blow-ups in hypothesis size (Koopmann, 2021). Another direction concerns extending from ABox to TBox repair-style abduction while retaining stronger relevance notions such as connection minimality (Haifani et al., 2022).

Taken together, the literature presents abduction under repair semantics as a distinct explanatory paradigm for inconsistent knowledge bases: the goal is not merely to make an observation derivable, but to do so in a way that remains meaningful across repaired views of inconsistent data, and, when desired, without introducing new conflicts or by minimizing the conflicts introduced (Haak et al., 2 May 2026).