In knowledge representation, ontologies serve as the structured blueprints of domain knowledge, mapping out concepts, roles, and the logical relationships between them. However, as ontologies mature, they frequently suffer from "logical bloat"—accumulating massive layers of complexity, redundant rules, and highly specialized vocabulary that make automated reasoning computationally expensive and human management error-prone.
When an application only requires a small fraction of a massive ontology's vocabulary, we face a fundamental challenge: how do we shrink the ontology to focus only on the relevant terms without breaking the underlying logic?
The Failure of Simple Deletion
The most intuitive approach to simplifying an ontology is syntactic deletion: simply removing any axioms that mention the unwanted concepts or roles. However, this naive approach routinely breaks the reasoning chains required for accurate query answering.
Consider a simple medical ontology containing two axioms:
- $\text{Infarction} \sqsubseteq \text{TissueDamage}$
- $\text{TissueDamage} \sqsubseteq \text{MedicalCondition}$
If an application does not care about the intermediate concept $\text{TissueDamage}$, simply deleting it leaves us with only $\text{Infarction} \sqsubseteq \text{TissueDamage}$ (which is invalid because the term is gone) or nothing at all. Crucially, we lose the implicit, mediated relationship that an $\text{Infarction}$ is a $\text{MedicalCondition}$. As noted by [Zhe Wang] et al. in Forgetting and Unfolding for Existential Rules, removing intermediate predicates without a formal mechanism destroys the deductive pathways necessary to answer external queries correctly.
The Intuition of Logical Forgetting
Logical forgetting (also known as uniform interpolation) solves this problem by shifting the focus from syntactic cleanup to semantic preservation. As described by [Frank van Harmelen] et al. in Preface - Special issue on commonsense reasoning for the semantic web, logical forgetting is a formal operation that eliminates a target set of terms (the "forgotten" signature) from a knowledge base while mathematically guaranteeing that all logical consequences formulated using the remaining vocabulary are perfectly preserved.
An intuitive way to visualize logical forgetting is as a geometric projection of a logical space.
Imagine a three-dimensional shape casting a shadow onto a two-dimensional plane. The shadow represents a simplification—it drops one entire dimension (the forgotten term)—yet it accurately preserves the profile and boundaries of the original shape along the remaining two dimensions. In description logics, the uniform interpolant is that perfect logical shadow. It is a new, often more concise set of axioms over the restricted vocabulary that yields the exact same answers to any query as the original, bloated ontology.
How It Works: Mechanisms of Elimination
To compute this logical shadow without losing information, researchers rely on two primary algorithmic paradigms: quantifier elimination and knowledge compilation.
1. Resolution and Quantifier Elimination
From a semantic standpoint, forgetting a concept $B$ is equivalent to existential second-order quantifier elimination ($\exists B . \mathcal{O}$). To achieve this algorithmically, calculi often utilize resolution. If we resolve every clause containing a positive occurrence of $B$ against every clause containing a negative occurrence of $B$, the resulting resolvents capture all the implicit logical dependencies that $B$ mediated. Once these resolvents are added to the ontology, all axioms containing $B$ can be safely discarded.
For highly expressive description logics, [Yizheng Zhao] et al. in Forgetting Concept and Role Symbols in ALCOIHµ(∇, ⊓)-Ontologies demonstrate that an Ackermann-based calculus can be used to systematically eliminate both concept and role symbols. By transforming axioms into a clausal form and applying Ackermann's lemma, the algorithm isolates the target symbols and replaces them with their logical definitions relative to the remaining vocabulary, bypassing the need for manual rule engineering.
2. Knowledge Compilation
An alternative paradigm compiles the ontology into a target normal form where forgetting becomes computationally trivial. [Meghyn Bienvenu] introduces this approach for description logics in Prime Implicate Normal Form for ALC Concepts.
By pre-processing complex logical expressions into a Prime Implicate Normal Form (PINF)—a recursive, redundancy-free structure—the structural dependencies of all terms are explicitly laid bare. Once an ontology is compiled into PINF, forgetting specific atomic terms or roles ceases to be a hard reasoning problem; instead, it can be executed via a simple, linear-time structural traversal that strips out the target terms while leaving the surrounding compiled logic intact.
The Challenge of Expressivity Leaks
If uniform interpolation is so powerful, why isn't it used universally? The core difficulty lies in a phenomenon known as an expressivity leak.
When you eliminate intermediate terms, the remaining logic must sometimes express highly complex, recursive relationships that were previously simplified by the forgotten terms. [Carsten Lutz] and [Frank Wolter] established the theoretical boundaries of this problem in Foundations for Uniform Interpolation and Forgetting in Expressive Description Logics, proving that deciding whether a finite uniform interpolant exists in the standard description logic $\mathcal{ALC}$ is 2-EXPTIME-complete. When it does exist, the resulting ontology can explode to triple-exponential size. In worse cases, an exact representation might require advanced logical constructs—like fixpoints or nominals—that the original logic did not possess.
To make forgetting practical, researchers employ two main strategies:
- Logic-Preserving Constraints: To prevent simplified ontologies from becoming incompatible with standard reasoning tools, [Yizheng Zhao] et al. design algorithms in A Practical Approach to Forgetting in Description Logics with Nominals that strictly guarantee the output remains within the original description logic fragment, avoiding the introduction of complex fixpoints.
- Complete Approximations: When a perfect, finite uniform interpolant is mathematically impossible to represent in the target logic, we can approximate it. [Kewen Wang] et al. in Concept and Role Forgetting in ALC{\mathcal {ALC}} Ontologies and [Anneke Bötcher] et al. in Complete Approximations of Horn DL Ontologies propose using sequences of finite, depth-bounded approximations. These approximations are mathematically complete up to a specified logical depth, allowing users to trade off absolute theoretical perfection for practical tractability.
From Schemas to Real-World Data
While early research focused heavily on forgetting terms within terminological schemas (TBoxes), real-world applications must manage instance data (ABoxes) and rule-based systems.
Applying forgetting to factual data assertions introduces significant complexity. As [Patrick Koopmann] and [Renate A. Schmidt] demonstrate in Uniform Interpolation and Forgetting for ALC Ontologies with ABoxes, eliminating concepts at the data layer can trigger a combinatorial explosion of disjunctive facts to account for every possible individual relationship that the forgotten concept originally organized.
When scaled to database-like frameworks using existential rules, the authors must ensure that query answering remains invariant. As shown by [Zhe Wang] et al. in Forgetting and Unfolding for Existential Rules, this is achieved through query-preserving predicate elimination, which guarantees that any Boolean Conjunctive Query evaluated over the simplified, forgotten ontology returns the exact same results as the original dataset.
Practical Applications
By providing a mathematically rigorous way to shrink signatures without losing deductive power, logical forgetting serves as a foundational utility across several domains:
- Deductive Module Extraction: Instead of forcing a system to reason over a massive biomedical ontology, uniform interpolation can extract a highly compact, self-contained "deductive module" focused entirely on a target signature. [Hui Yang] et al. demonstrate in Efficient Computation of General Modules for ALC Ontologies and Efficient Extraction of EL-Ontology Deductive Modules that this technique yields modules significantly smaller and faster to process than traditional syntax-based extraction methods.
- Ontology Reshaping: In industrial environments, general-purpose ontologies are often too abstract for specific deployments. [Dongzhuoran Zhou] et al. in Towards Ontology Reshaping for KG Generation with User-in-the-Loop: Applied to Bosch Welding utilize forgetting to streamline complex schemas into data-specific knowledge graphs, cutting graph depth in half and accelerating generation speeds by up to 8x.
- Conflict Resolution and Updates: When updating an ontology with new information, contradictions frequently arise. Rather than deleting entire axioms to resolve a conflict, [Guilin Qi] and [Jianfeng Du] demonstrate in Model-based Revision Operators for Terminologies in Description Logics that concept forgetting can act as a fine-grained surgical tool. It systematically removes only the problematic terms causing the contradiction, allowing the rest of the axiom's logical implications to remain intact. Similarly, [Yan Zhang 0003] and [Norman Y. Foo] in A Unified Framework for Representing Logic Program Updates show that strong and weak forgetting operate as universal primitives capable of simulating diverse logic program transformation and update strategies.
Key Foundational Work
Building a Large Knowledge Base for a Natural Language System (1984) — contributed 1 concept elements to this direction.
Assessing the Maintainability of XCON-in-RlME: Coping with the Problems of a VERY Large Rule-Base (1987) — contributed 1 concept elements to this direction.