Distance-based Paraconsistent Semantics

• Distance-based paraconsistent semantics is defined by selecting models that minimize the distance from classical, consistent representations using metrics like Hamming distance.
• The approach preserves robust, non-explosive reasoning by supporting relevant entailments even when the knowledge base contains inconsistencies.
• It extends to description logics and repair-based systems, enabling efficient, fine-grained, and personalized query answering in complex, inconsistent settings.

Distance-based paraconsistent semantics constitute a class of approaches for non-explosive, relevance-sensitive reasoning in the presence of inconsistency, unified by the principle of selecting models (or repairs) that are “closest” to the classical (consistent) ideal according to a well-defined metric. These methods aim to preserve as much information as possible from an inconsistent knowledge base (KB), supporting meaningful entailment without trivializing inference. In contrast to purely syntactic repair or model-elimination approaches, distance-based techniques introduce explicit measures of proximity between representations (types, models, or repairs), yielding fine-grained and semantically driven trade-offs among conflicting axioms or facts.

1. Distance-based Paraconsistent Semantics in Description Logic

Distance-based paraconsistent semantics were formally introduced for DL-Lite by Zhang et al. through the notion of features—finite, coherent mini-models constructed from types and Herbrand sets (Zhang et al., 2013). The key elements are:

• Types and Features: Every type is a maximal consistent set of basic concepts within the signature; features pair collections of such types with finite Herbrand sets, encoding local assertions.
• Distance Measures: The core innovation is the definition of pseudo-metric functions on types (e.g., Hamming distance or drastic distance), quantifying the deviation of a given type from satisfying all TBox constraints.
• Model Construction: For a TBox, minimal model-types are those that minimize the aggregated distance from the set of type-constraints induced by the TBox's inclusions. For full KBs (TBox plus ABox), profiles for individuals are computed and minimal types chosen similarly, ensuring role coherence and Herbrand maximality.
• Paraconsistent Entailment: An axiom φ is distance-based paraconsistently entailed, denoted $𝓚 ⊨_{d,f} φ$, if all df-minimal model features satisfy φ.

This semantic framework guarantees non-explosiveness (no contradictions are entailed from inconsistency), preserves classical entailments in the absence of inconsistency, and ensures that consequences remain relevant and non-trivial, even in globally inconsistent settings.

2. Logical Properties and Theoretical Guarantees

Distance-based paraconsistent entailment ($⊨_{d,f}$) for DL-Lite possesses a spectrum of desirable logical properties (Zhang et al., 2013):

• Paraconsistency: $⊨_{d,f}$ never explodes; inconsistent KBs never entail arbitrary or contradictory statements.
• Consistency Preservation: When the KB is consistent, this entailment coincides with classical logic.
• Relevance and Non-monotonicity: Inferences depend only on the relevant fragment of the vocabulary and, due to minimality, entailment is non-monotonic (additional knowledge can both create and eliminate consequences).
• Cautious Monotonicity/Splitting: For KBs with disjoint signatures, classical consequences from a consistent fragment are preserved.
• Closure Consistency: The set of distance-entailed inclusions (ABox/TBox) is itself consistent.
• TBox Preservation: For inclusions, entailment with data equals entailment at the type level alone.

Comparison with alternative approaches highlights key distinctions: unlike four-valued or quasi-classical logics, the semantics operate classically on chosen models; unlike syntax-based repair, the process is determined semantically via distances; and by working with features, infinite models are avoided, affording practical computation.

3. Distance-based Semantics for Three-Valued Paraconsistent DLs

A related instantiation in more expressive DLs such as $\mathcal{ALC}$ is developed around conflict-minimal three-valued interpretations (Qiao et al., 2014). Here, interpretations assign both positive and negative extensions to each atom—a conflict occurs when both hold nontrivially. The principal distance notion is the cardinality of conflict assignments:

• Model Selection: Among all three-valued models, select those minimizing the number of conflicts.
• Entailment: An assertion $\varphi$ is entailed in this minimal-conflict sense, written $Σ ⊨_{<ₙ} \varphi$, if every conflict-minimal model of $Σ$ supports $\varphi$ as “at least true.”

A notable consequence is the induced non-monotonicity: as knowledge bases are extended, conflict-minimality may force further conflicts, removing previously held consequences. This behavior contrasts classical monotonicity and is diagnostically relevant for applications needing cautious expansion of inconsistent KBs.

Assumption-based argumentation frameworks (ABA) provide a precise computational counterpart: branches in semantic tableaux are closed under assumptions preventing conflicts, and stable extensions of this ABA system correspond exactly to distance-based paraconsistent inferences in this setting.

4. Distance-based Semantics in Repair-based Ontology Approaches

Distance-based repair semantics are extended to ontology-based data access scenarios, where inconsistencies are treated by reasoning over repairs (maximal consistent subsets of facts) (Prouté et al., 2019). The innovations here are:

• Repair Distance Functions: For sets of facts, atom-level syntactic distances (e.g., Ramon–Bruynooghe metric employing least general generalisations) are lifted to sets via minimum-weight matchings.
• Clustering of Repairs: Instead of relying on the whole set of repairs, repairs are clustered by similarity; queries are then evaluated within these clusters using standard semantics (AR, IAR, ICR).
• Query Answering: Results become cluster-indexed: a query may be entailed in some clusters but not others, supporting personalized and fine-grained views of inconsistent entailment.

There is no introduction of new “d-certain” or “d-possible” inference modes; rather, the clustering partitions the entailment landscape, and global notions can be formulated via universal or existential quantification over cluster answers.

5. Illustrative Examples and Workflow

In DL-Lite, the canonical “Penguin problem” demonstrates the distance-based approach. An inconsistent KB containing both $Bird⊑Fly$ and $¬Fly(tweety)$, together with $⊨_{d,f}$0 and $⊨_{d,f}$1, would ascribe to $⊨_{d,f}$2 the consequences $⊨_{d,f}$3 and $⊨_{d,f}$4, while witholding $⊨_{d,f}$5 and $⊨_{d,f}$6. Thus, the semantics retrieve only the robust, intuitively correct consequences, avoiding trivialization (Zhang et al., 2013).

In the repair-based setting, for a knowledge base $⊨_{d,f}$7 with facts about two siblings and their attendance at different locations (day care, nanny, home), clustering repairs by distance results in blocks where some queries (such as whether “every baby gets ill”) are True in the first two clusters but False in the third, reflecting the diversity of maximal consistent perspectives under minimal change from the original data (Prouté et al., 2019).

6. Comparison with Other Paraconsistent Approaches and Discussion

Distance-based paraconsistent semantics depart from established approaches in multiple respects:

Approach	Model Selection	Explosiveness	Non-monotonic	Relevance Sensitive
Syntax-based repair (max. consistent sub-KB)	Syntactic selection of repairs	Non-explosive	Typically N	Often No
Four-valued/quasi-classical semantics	Truth value enrichment, all models considered	Non-explosive	No	Sometimes
Conflict-minimal three-valued models	Minimal conflicts in three-valued interpretations	Non-explosive	Yes	Yes
Distance-based (features, clusters)	Closest features/repairs by explicit metrics	Non-explosive	Yes	Yes

Distance-based semantics fuse semantic perspectives (closeness to classical models) with computational pragmatism (finite features, clustering), ensuring robust non-trivial entailment, efficient query answering (for DL-Lite and repairs), and a nuanced treatment of conflict that neither trivializes nor ignores inconsistency.

7. Practical Implications and Computational Considerations

Computation of distance-based paraconsistent entailment in DL-Lite is efficient due to the finiteness of the type space and direct manipulations of features (Zhang et al., 2013). For repair-based methods, computing all repairs can be exponential in the size of the fact set, but distance calculations and clustering (e.g., via spectral clustering on the similarity matrix) are polynomial in the number of repairs. Overall complexity remains tied to both the expressive power of the underlying ontology language and the combinatorics of repairs. Notably, no formal new complexity-theoretic lower or upper bounds are introduced for these frameworks beyond those of the underlying settings (Prouté et al., 2019).

Algorithmically, distance-based approaches require primitives for enumerating relevant model candidates (features, repairs), evaluating distances, and aggregating results via clustering and aggregation functions. In three-valued DLs, ABA frameworks are necessary for effective manipulation and justification of non-monotonic entailments (Qiao et al., 2014).

A plausible implication is that distance-based frameworks provide a systematic, semantics-grounded alternative for inconsistent reasoning in description logics and ontology data access, with particular strengths in fine-grained consequence selection, personalization of query answering, and avoidance of trivialization in the face of inconsistency.