The Ontology Revision

Yu Sun, Yuefei SuiIJCAI 2005

Abstract

In the field of knowledge representation, ontologies provide the structured hierarchies necessary for intelligent systems to categorize and reason about information. However, as new data emerges, these structures must be updated without compromising their logical consistency or structural integrity. This paper introduces a formal framework for ontology revision, proposing the "Z axiom system"—a hybrid approach that combines classical belief revision principles with rules for iterated updates. By focusing on ontology-base revision and maintaining tree-like subsumption relations, the authors provide a concrete operator capable of handling complex updates while ensuring minimal loss of existing knowledge.

Introduction

In modern artificial intelligence and knowledge management, an ontology is more than a simple collection of facts; it is a structured hierarchy of concepts and properties. For instance, an ontology might define that a "Golden Retriever" is a "Dog," and that "Dogs" possess the property of having fur. When new, potentially conflicting information arrives, the system must undergo ontology revision.

Traditional belief revision—the logic governing how an agent updates its knowledge—typically treats information as a "flat" set of sentences. Applying these standard methods to ontologies is problematic because ontologies are usually structured as taxonomies or trees. Simply deleting a contradictory statement, as standard belief revision might suggest, risks "breaking" the hierarchical tree or losing implicit information essential to the system's reasoning. This paper addresses the challenge of revising ontologies while satisfying three core principles: Success (incorporating new information), Consistency (avoiding self-contradiction), and Minimal Change (preserving as much of the original structure as possible).

The Z Axiom System and Methodology

The authors approach ontology revision through formal logic and axiomatic design, introducing the Z axiom system. This system consists of seven axioms (Z0–Z6) that define the criteria for a robust ontology revision. It represents a theoretical synthesis of two major frameworks:

  1. AGM Axioms: The established standard for one-shot belief revision.
  2. Darwiche-Pearl (DP) Axioms: The framework governing how revision should behave over multiple, iterated steps.

Ontology-Base Revision

Unlike "ontology-set" revision, which considers every possible logical deduction within a system, this work focuses on ontology-base revision. This approach modifies the actual explicit statements that constitute the ontology. This is a more practical approach for real-world systems where the explicit hierarchy is the primary object of management.

Structural Constraints and the Concrete Operator

A central assumption of this model is that the ontology maintains a tree structure under the subsumption ($⊑$) relation. To preserve this "tree-ness," the revision process does not merely delete facts. Instead, it employs a strategy of extraction and addition: it extracts specific consistent subsets of statements and adds new ones to ensure the hierarchy remains connected and valid.

To implement this, the authors define a specific mathematical operator ($∘$). This operator is designed to handle 18 distinct cases of interaction between new statements and existing knowledge, such as when a new statement contradicts a default property or challenges the core hierarchy.

Key Insights

Revision as Extraction and Addition

In standard logic, if statement A contradicts statement B, one is typically discarded. In the context of ontologies, the authors argue that preservation of structure requires a more surgical approach. By extracting valid components and adding necessary bridging statements, the system maintains structural integrity that a simple deletion would destroy.

Closed World Assumption (CWA)

The framework incorporates the Closed World Assumption, which posits that if a property is not explicitly stated or deducible, it is assumed to be false. This adds significant complexity to the revision process; adding a single new fact can trigger a cascade of contradictions with previously "assumed false" statements, requiring the revision operator to be highly sensitive to implicit negations.

Atomic Revision

To maintain granular control over the rewiring of the hierarchy, the authors focus on atomic revision. This involves updating the ontology with single relationships (e.g., $C ⊑ D$). This simplification allows for a more precise mathematical treatment of how individual nodes in the taxonomy are moved or redefined.

Results and Theoretical Proofs

The paper provides two primary theoretical contributions:

  • Theorem 1: The authors prove that the Z axiom system successfully satisfies the three pillars of belief revision: success, consistency, and minimal change.
  • Theorem 2: They demonstrate that the proposed concrete revision operator ($∘$) is a valid implementation of the Z axiom system, providing a functional tool for developers of knowledge-based systems.

Furthermore, by integrating the DP axioms, the framework is shown to be robust for iterated revision. This ensures that the system remains stable over a sequence of many updates, a requirement for AI systems that learn and evolve over time.

Limitations and Future Work

While providing a foundational framework, the authors identify several areas for expansion:

  • Logic Complexity: The current model utilizes relatively simple logical relations. Future research will focus on integrating more complex Description Logics (DL) and first-order logic.
  • Complex Constraints: Currently, the system handles atomic statements. Real-world updates often involve complex concepts (e.g., "A is a B and not a C"), which require more sophisticated handling.
  • Property Integration: The relationship between different types of statements—such as how strict inheritance interacts with default inheritance—remains simplified and warrants deeper integration into the revision logic.

Summary updated

BibTeX
@inproceedings{sun2005ontology,
  author = {Yu Sun and Yuefei Sui},
  title = {{The Ontology Revision}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2005},
}
Figure