Rational Closure For All Description Logics (Extended Abstract)

Piero A. BonattiIJCAI 2020 · DOI

TL;DR

This paper tackles the challenge of extending "rational closure," a nonmonotonic reasoning mechanism for description logics (DLs), to more expressive DLs like OWL2. The authors demonstrate that standard rational closure fails for these expressive logics and introduce "stable rational closure," a refined approach that overcomes these limitations, though it reintroduces the possibility of non-unique or non-existent deductive closures.

What Problem Does This Paper Address?

Modern applications like biomedical ontologies and semantic web policies increasingly rely on description logics (DLs) to represent knowledge. These applications often require a way to handle "default" or "defeasible" knowledge – statements that are generally true but can be overridden by more specific information. For example, "eukaryotic cells normally have a nucleus," but "mammalian red blood cells normally do not have a nucleus."

Rational Closure (RC) is a well-studied nonmonotonic semantics designed to handle such defeasible inclusions (DIs). It aims to provide a unique and logically sound way to reason with default knowledge, adhering to desirable properties known as the KLM postulates. However, existing RC semantics are limited to simpler DLs and cannot be directly applied to more expressive logics like SROIQ(D), which forms the basis of OWL2. This limitation arises because the standard RC semantics relies on a property called the "disjoint model union property" (DMUP), which is not satisfied by expressive DLs that include features like nominals (specific individuals) or universal roles. This paper addresses this gap by showing the limitations of standard RC and proposing a new, more general version.

Methods

The authors first prove that standard rational closure, with its model-theoretic semantics based on canonical models, cannot be extended to DLs that do not satisfy the disjoint model union property (DMUP). This is a crucial theoretical result showing the inherent limitations of the existing framework.

To overcome this, they introduce stable rational closure (stable RC). The core idea is to modify the definition of the "ranking" used in RC. Instead of relying on a ranking derived from subsets of the knowledge base (which can lose information), stable RC uses a ranking function that considers the entire knowledge base at each step. This is achieved by setting up and solving a "meta-level equation" that captures the requirements for these new ranking functions.

Stable RC defines a new notion of "upward-closed models" which play a similar role to canonical models in standard RC. Crucially, the existence of these upward-closed models does not depend on the DMUP, making stable RC applicable to a wider range of DLs. They also introduce new "exceptionality criteria" for ranking functions that are essential for the model-theoretic characterization of stable RC.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • DMUP is the Bottleneck: The paper clearly demonstrates that the disjoint model union property (DMUP) is the fundamental reason why standard rational closure fails for expressive DLs.
  • Stable RC: A Generalization: Stable RC is presented as a natural extension of RC. It coincides with RC when the DMUP holds, and it preserves desirable computational properties.
  • The Trade-off: While stable RC overcomes the DMUP limitation, it reintroduces a drawback common to many nonmonotonic logics: deductive closures may not exist or may not be unique. This means that reasoning with stable RC might sometimes lead to ambiguity or inconsistency.
  • Ranking Matters: The paper highlights the critical role of the ranking function in defining nonmonotonic semantics and shows how modifying this function can address fundamental limitations.
  • Upward-Closed Models: The introduction of upward-closed models provides a new semantic foundation for defeasible reasoning that is independent of the DMUP.

Key Results

  • Theorem 1: Proves that standard RC is not complete for DLs lacking the DMUP, meaning some valid inferences might be missed.
  • Stable RC Definition: Introduces a refined definition of rational closure that does not depend on the DMUP.
  • Theorem 2: Establishes a model-theoretic characterization of stable RC using upward-closed models.
  • Theorem 3: Shows that stable RC satisfies the KLM postulates, a set of desirable properties for nonmonotonic reasoning.
  • Theorem 4: Demonstrates that stable RC does not increase the computational complexity of reasoning for many DLs, preserving tractability.
  • Theorem 5: Shows that when the DMUP holds, stable RC is equivalent to standard RC.
  • Theorem 6: Reveals that for DLs lacking DMUP, knowledge bases may have no stable rankings or multiple stable rankings, leading to potential non-uniqueness of conclusions.

Limitations and Future Work

The authors acknowledge that a significant drawback of stable RC is the reintroduction of non-unique or non-existent deductive closures, similar to other nonmonotonic logics. They also mention inherited drawbacks from RC, such as default properties not applying to role fillers and the "inheritance blocking" phenomenon where one overriding can prevent the inheritance of other default properties.

Future work suggested includes completing and integrating existing research efforts to make stable RC more practical, particularly by addressing the generality problem and defining nonmonotonic inferences more concretely. The authors also point to DLN (a competing logic) as an appealing alternative that avoids some of these drawbacks, though it has its own unique way of handling unresolved conflicts.

Summary updated

BibTeX
@inproceedings{bonatti2020rational,
  author = {Piero A. Bonatti},
  title = {{Rational Closure For All Description Logics (Extended Abstract)}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/700},
}
Figure