A Terminological Logic with Defaults: A Definition and an Application

Lin Padgham, Tingting Zhang

Abstract

This paper introduces Default Terminological Logic (DTL), a framework that bridges the gap between rigid terminological hierarchies and flexible default inheritance systems. By bifurcating concepts into "core" (necessary) and "default" (typical) components, DTL enables more nuanced knowledge representation. The authors demonstrate the practical utility of DTL through a medical diagnosis system that achieves specialist-level performance on real-world clinical protocols.

Introduction

In the early 1990s, Knowledge Representation (KR) faced a significant divide. Terminological logics, such as KL-ONE, were adept at defining strict, necessary hierarchies but struggled to accommodate exceptions. Conversely, default inheritance systems could handle "usually true" statements but lacked the formal rigor of terminological systems. This dichotomy posed a major hurdle for real-world applications like medicine, where a disease might have both "necessary" symptoms (essential for diagnosis) and "typical" symptoms (present in most, but not all, cases). Traditional logics forced a choice between brittleness and imprecision.

The DTL Framework

The authors propose DTL (Default Terminological Logic) to address this gap. The central innovation is the Core/Default Split: every concept in the hierarchy is represented by two distinct nodes:

  1. Core node ($C_c$): Contains properties that are strictly necessary.
  2. Default node ($C_d$): Contains properties that are typically true.

A semantic constraint ensures that the core is always subsumed by the default. This structure allows for "Default Subsumption," a flexible relationship where a concept $X$ default-subsumes $Y$ if $X$ is default-reachable from $Y$. This allows for a more flexible classification of individuals who might not meet every strict criterion of a category.

Figure 1: Full vs. simplified representation of core and default nodes in DTL, illustrating the default-subsumption relationship.

To manage potential conflicts in default logic, DTL employs skeptical inheritance. The system only accepts a default conclusion if no competing, equally valid arguments exist against it. Furthermore, the authors intentionally limited the language's complexity (e.g., excluding complex relational constructs) to maintain computational tractability while providing the expressivity needed for medical domains.

Medical Diagnosis Application

The practical power of DTL was tested by building an interactive diagnosis engine. The system frames medical diagnosis as a "minimal cover" problem: the goal is to find the smallest set of disease concepts that collectively explain all observed patient symptoms.

The knowledge base was constructed using data from standard medical textbooks, such as WHO manuals and Harrison’s Principles of Internal Medicine, focusing on genito-urinary diseases. The system operates in a non-monotonic loop, performing an initial "approximate default classification" and then using information theory to identify the most effective follow-up questions or tests.

Figure 3: An example of the diagnostic process, showing how patient symptoms are mapped to disease concepts to find a minimal cover.

Results and Evaluation

The DTL system was evaluated against 63 real patient protocols collected by Dr. Toomas Timpka. To establish a gold standard, these cases were also reviewed by a General Practitioner (GP) and nine specialists. The results demonstrated remarkable robustness:

  • Total Agreement: In cases where all 10 human doctors agreed, the DTL system matched their diagnosis 100% of the time.
  • Consensus Matching: In cases where doctors initially disagreed but reached a consensus, the system matched the consensus in 91% of instances (20 out of 22 cases).
  • Specialist-Level Accuracy: The system's performance was on par with human specialists, whose individual match rates with the consensus ranged from 41% to 91%.
  • Efficiency: The diagnostic engine typically narrowed down the correct taxonomic path within just three iterations of its question-and-answer loop.

Conclusion

DTL successfully integrates default reasoning into taxonomic hierarchies, providing a formal yet flexible tool for complex domains. While the current implementation lacks relational "roles" and uses a simplified "approximate" classification for individuals, its performance in medical diagnosis proves the viability of the core/default split. Future work may explore more complex fix-point reasoning and investigate whether the observed pattern of core-to-core inheritance is a universal feature of human knowledge organization.

Summary updated

BibTeX
@inproceedings{padgham1993terminological,
  author = {Lin Padgham and Tingting Zhang},
  title = {{A Terminological Logic with Defaults: A Definition and an Application}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1993},
}
Figure