Logic-Based Ontology Modularization for Scalable Reasoning

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 255 papers91 conceptsfrozen 2026-10-04

Modern knowledge bases, such as the medical terminology SNOMED CT or the National Cancer Institute Thesaurus (NCIt), are massive. They contain hundreds of thousands of axioms describing complex relationships between diseases, anatomy, and treatments. However, a researcher studying diabetic retinopathy doesn't need the entire medical universe; they only need the logic relevant to the eye and diabetes.

Ontology modularization is the formal process of carving out these small, manageable subsets. By extracting a "module," we create a fragment that is significantly smaller than the original but remains logically identical for a specific set of terms.

The Curse of the Monolith

The primary motivation for modularization is computational. As P. Pandurang Nayak demonstrated in Representing Multiple Theories, reasoning over a large, monolithic knowledge base is often intractable. By partitioning a system into localized contexts, the complexity of reasoning drops from being dependent on the total vocabulary to the size of the largest individual module. This shift can prevent the satisfiability problem from escalating from NP-complete to PSPACE-complete.

In practical terms, Mercedes Argüello Casteleiro and colleagues showed in Experiments to Create Ontology-based Disease Models that extracting a locality-based module can reduce the computational overhead of a clinical ontology by over 99%. This reduction allows complex formal logic to run in real-time clinical applications where processing the full terminology would be impossible.

Defining the Boundary: The Signature

To extract a module, we first define a signature: the specific set of concepts and roles we care about. The goal is to find a subset of the ontology that is "inseparable" from the original regarding that signature.

Boris Konev, Frank Wolter, and others explored this in Conjunctive Query Inseparability of OWL 2 QL TBoxes. They framed the problem as ensuring that for any query using only the target vocabulary, the module and the original ontology return the exact same answers.

How Extraction Works: Locality and Datalog

How do we know which axioms to keep? The most common approach uses syntactic locality. This technique, detailed by Bernardo Cuenca Grau and Ulrike Sattler in Safe and Economic Re-Use of Ontologies, uses syntactic conditions to identify axioms that "could" affect the meaning of the signature. It is a fast, polynomial-time approximation of the logically minimal module.

A more precise, "semantic" approach was introduced by Ana Armas Romero and Ian Horrocks in Ontology Module Extraction via Datalog Reasoning. Their method works by:

  1. Skolemization: Replacing complex existential quantifiers with constants to simplify the logic.
  2. Datalog Reduction: Transforming the complex Description Logic axioms into a Datalog program.
  3. Materialization: Running a Datalog reasoner to see which axioms are actually triggered to reach conclusions about the signature.

This "pay-as-you-go" approach, also seen in the PAGOdA system, allows for much smaller modules than standard syntactic methods because it ignores axioms that are syntactically related but logically redundant for the specific task.

Forgetting and Views

Sometimes, instead of picking what to keep, it is easier to decide what to "forget." Deductive forgetting (or uniform interpolation) is a technique where specific concepts are removed while their logical consequences are preserved among the remaining terms.

Tools like MEX and CEX, developed by Carsten Lutz and Dirk Walther, use these principles to perform "logical diffs." As described in Logical Difference and Module Extraction with CEX and MEX, this allows developers to see exactly how the meaning of an ontology has changed between versions, rather than just seeing which lines of text were edited.

Applications in the Wild

Modularization isn't just a theoretical trick; it is a prerequisite for many modern AI tasks:

  • Data Integration: Shqiponja Ahmetaj and Phokion G. Kolaitis showed in Ontology-Enriched Query Answering on Relational Databases that pruning ontologies to match a database's schema can reduce size by 90%, making query answering feasible.
  • Conflict Resolution: When merging ontologies, contradictions are common. Guilin Qi used locality-based modules in A Conflict-Based Operator for Mapping Revision to isolate the specific axioms causing a conflict, ignoring the rest of the massive knowledge base during the repair process.
  • Distributed Reasoning: In multi-agent systems, George Vouros proposed the mONTul method, which treats modularization as a constraint satisfaction problem to optimally distribute knowledge across different agents while maintaining global consistency.

By breaking the monolith, modularization transforms "Big Knowledge" from a computational burden into a flexible, reusable resource.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 55 papers
  2. Version 1 · 2026-10-04 · 55 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_ontology_modularization_0739b021,
  title = {ontology modularization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/ontology-modularization/rdr_0739b021daed4d3f89ae36f337eb4861},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_logic_based_ontology_modularization_for__0739b021_v1,
  title = {Logic-Based Ontology Modularization for Scalable Reasoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/logic-based-ontology-modularization-for-scalable-reasoning/rdr_0739b021daed4d3f89ae36f337eb4861/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}