Tractable Reasoning and Conjunctive Query Answering for Defeasible DL-Lite under Rational Closure

What Happened

A new arXiv preprint (2606.24279v1) tackles a foundational problem in knowledge representation: how to make Description Logics (DLs) — the formal backbone of ontologies and semantic reasoning — handle defeasible or "typically true" information without collapsing into logical inconsistency. The authors focus on applying Rational Closure (RC), a well-established non-monotonic reasoning framework, to the core and Horn fragments of DL-Lite, a lightweight but widely used family of description logics. Specifically, they investigate tractable reasoning and conjunctive query answering under this defeasible regime.

The key technical contribution is showing that for DL-Lite core and Horn, rational closure can be computed efficiently — in polynomial time — and that conjunctive query answering remains tractable. This is non-trivial because adding defeasible axioms (e.g., "typically, birds fly") to classical DLs often leads to exponential blow-ups or undecidability. The paper provides algorithms and complexity bounds that establish a sweet spot: expressive enough for realistic defeasible ontologies, yet computationally manageable.

Why It Matters

This work addresses a long-standing tension in AI knowledge representation. Classical DLs like OWL 2 are excellent for crisp, monotonic reasoning — if you know all exceptions, they work perfectly. But real-world knowledge is rife with exceptions: "Most employees are full-time" doesn't mean every employee is. Without defeasible reasoning, ontologies either become brittle (requiring explicit exception lists) or lose nuance.

Rational closure has been a theoretical favorite for decades because it preserves the "specificity principle" — more specific information overrides general defaults. However, its computational cost has limited practical adoption. By proving tractability for DL-Lite core and Horn — the same fragments used in many OWL 2 profiles and biomedical ontologies — this paper opens the door to integrating defeasible reasoning into production systems without sacrificing performance.

For AI practitioners building knowledge graphs, decision support systems, or semantic search engines, this means you can now model "typical" relationships (e.g., "typically, a paper has an author") and still answer queries like "find all papers co-authored by X" efficiently, even when some papers are authorless (e.g., anonymous submissions). The formalism handles exceptions gracefully without requiring manual curation.

Implications for AI Practitioners

First, ontology engineers can begin designing defeasible schemas for DL-Lite-based systems with confidence that query answering will remain polynomial. This is particularly relevant for biomedical and enterprise ontologies where default knowledge is pervasive (e.g., "typically, a gene codes for a protein").

Second, the results provide a clear boundary: if your application needs defeasible reasoning, stick to DL-Lite core or Horn fragments. Moving to more expressive DLs (e.g., with full role hierarchies or disjunction) may still incur exponential costs under rational closure.

Third, this work complements recent advances in neuro-symbolic AI. Tractable defeasible DLs could serve as a symbolic "safety net" for neural models that generate or interpret knowledge, ensuring logical consistency even when neural outputs are noisy.

Key Takeaways

• Tractable defeasible reasoning is now possible for DL-Lite core and Horn under rational closure, with polynomial-time algorithms for both satisfiability and conjunctive query answering.
• Practical ontologies can model default knowledge without manual exception lists, enabling more natural knowledge representation in biomedical, enterprise, and semantic web applications.
• The DL-Lite family remains the sweet spot for combining expressivity with computational efficiency — even under non-monotonic extensions.
• AI practitioners should watch for implementation releases; the theoretical results here are ripe for integration into OWL reasoners like Pellet, HermiT, or ELK.