In classical logic, a single contradiction can be catastrophic. If a knowledge base contains both "all birds fly" and "penguins are birds that do not fly," a standard reasoner might conclude that everything is true—a state known as logical explosion. To prevent this, researchers have traditionally relied on belief revision, a process of resolving conflicts by discarding old information to make room for the new.
However, traditional revision is often a blunt instrument. As Salem Benferhat, Didier Dubois, and Henri Prade argued in How to Infer from Inconsistent Beliefs without Revising?, deleting entire formulas to restore consistency results in the permanent loss of potentially valuable information.
The emerging consensus across three decades of research is a shift toward logical forgetting (or uniform interpolation). This mechanism allows systems to surgically remove specific concepts or roles while preserving all logical implications between the remaining terms, enabling knowledge bases to evolve without losing their inferential power.
The Problem with Blunt Deletion
When a knowledge base is updated, the goal is "minimal change"—retaining as much prior knowledge as possible while incorporating new data. Hirofumi Katsuno and Alberto O. Mendelzon established the foundational requirements for these "rational" changes in A Unified View of Propositional Knowledge Base Updates.
The difficulty lies in the difference between explicit and implicit knowledge. If you delete an explicit axiom to resolve a conflict, you might inadvertently destroy implicit consequences that were not in conflict. Maurice Pagnucco and Zhiqiang Zhuang note in Belief Contraction in the Description Logic EL that targeting only explicit statements risks losing valuable logical entailments.
This leads to two common failures:
- The Drowning Effect: In prioritized systems, a contradiction at a high-certainty level can cause the system to ignore all lower-certainty information, even if it is unrelated to the conflict Guilin Qi in A Semantic Approach for Iterated Revision in Possibilistic Logic.
- Structural Ruin: Purely syntactic updates can render an ontology unreadable or break the intended schema structure by treating human-modeled axioms and derived facts as equally expendable Bernardo Cuenca Grau et al. in How to Contract Ontologies.
Forgetting as a Surgical Tool
Logical forgetting offers a more refined alternative. Instead of deleting an entire axiom, forgetting allows a system to "marginalize" a specific term.
Imagine an ontology describing medical treatments. If a specific drug is discontinued due to privacy concerns or new side-effect data, you don't want to delete every research paper or patient record associated with it. Instead, you want to "forget" the drug name while keeping all the logical relationships between the symptoms it treated and the outcomes observed.
Kewen Wang and colleagues have demonstrated that this process can be applied to various formalisms, from Logic Programming to Existential Rules. In Concept and Role Forgetting in ALC Ontologies, the authors show that forgetting acts as a bridge: it allows us to calculate what a knowledge base would look like if a certain concept were removed, without manually rewriting the entire logic.
Bridging Models and Syntax
One of the most powerful applications of forgetting is in model-based revision. In this approach, we don't look at the text of the axioms; we look at the "worlds" the axioms allow.
As Guilin Qi and Jianfeng Du explain in Model-based Revision Operators for Terminologies in Description Logics, forgetting serves as the computational engine for this. By systematically removing problematic terms, the system can construct a new, consistent terminology that resolves contradictions without the "all-or-nothing" loss of entire axioms.
This is particularly useful in distributed ontologies. When merging data from different sources, contradictions are inevitable. Rather than discarding an entire local ontology, Guilin Qi and Pascal Hitzler suggest using forgetting to selectively ignore the specific concept names causing the clash, restoring consistency while preserving the rest of the shared knowledge A Forgetting-based Approach for Reasoning with Inconsistent Distributed Ontologies.
The Limits of Preservation
While logical forgetting is powerful, it is computationally demanding. Carsten Lutz and Frank Wolter proved in Foundations for Uniform Interpolation and Forgetting in Expressive Description Logics that for expressive logics like $\mathcal{ALC}$, the resulting simplified ontology can be triple-exponentially larger than the original.
Furthermore, there is the risk of expressivity leaks. Sometimes, forgetting a simple concept in a basic logic requires a more complex logic to represent the remaining relationships. To combat this, researchers like Yizheng Zhao have developed practical tools like UI-FAME that use specialized calculi to keep the output within the original logic fragment, ensuring the results remain compatible with standard reasoning tools.
Summary
The shift from deletion-based revision to forgetting-based evolution represents a move toward more resilient AI systems. By treating knowledge not as a collection of fragile strings, but as a web of logical consequences, researchers are creating ways for machines to "change their minds" or "simplify their views" without losing the hard-won insights buried in their data.