A Terminological Logic with Defaults: A Definition and an Application

TL;DR

This paper introduces DTL, a terminological logic that extends traditional systems by incorporating default reasoning. This allows for more flexible knowledge representation, particularly in domains where information is incomplete or uncertain, and demonstrates its utility through a successful medical diagnosis application.

What Problem Does This Paper Address?

Traditional terminological logic (TL) systems, while powerful for representing hierarchical knowledge and performing classification, often struggle with real-world scenarios where information is not absolute. They typically rely on strict inheritance, meaning a property inherited by a concept is always true. This is problematic for domains like medicine, where symptoms might be typical but not strictly guaranteed, or where knowledge is incomplete. Existing default reasoning approaches have been theoretically explored but have seen limited practical application. This paper aims to bridge this gap by defining a practical terminological logic with defaults and demonstrating its effectiveness in a real-world application.

Methods

The core contribution is the definition of DTL (Default Terminological Logic), a language that extends standard TLs by including default constructs. Key features include:

  • Core and Default Nodes: Concepts are represented by both a "core" node (necessary properties) and a "default" node (typical properties). A crucial semantic constraint is that the core is always subsumed by the default.
  • Default Subsumption: A new definition of subsumption is introduced, called "default subsumption." This relies on a "default reachability" relationship, which is determined by a skeptical default inheritance reasoner. This allows for reasoning about typical properties even when strict inheritance might not apply.
  • Classification Algorithm: The paper presents an algorithm for classifying new concepts into the DTL taxonomy. This algorithm adapts standard classification by using default subsumption for default concepts and strict subsumption for core concepts, ensuring that necessary properties are always guaranteed.
  • Default Classification of Individuals: For reasoning about individuals (e.g., diagnosing a patient), the system performs an "approximate default classification." This process prioritizes default subsumption to explore the hierarchy more deeply and identify probable diagnoses, even with incomplete information.

Datasets

The paper describes an evaluation using:

  • 63 Patient Protocols: These were collected from a general practitioner at a community health center and included symptoms, signs, and lab test results. The GP's diagnosis was also recorded. These protocols were then used to evaluate the system's diagnostic accuracy against medical doctors.

Key Insights and Ideas

  • Defaults are Crucial for Real-World Applications: The paper strongly argues that for many practical domains, especially medicine, strict inheritance is insufficient. Defaults are a necessary construct for representing typicality and handling uncertainty.
  • A Practical Trade-off: DTL offers a balance between expressivity and computational manageability. It's simpler than many existing TLs but includes essential features like defaults and features with predicate values (though the latter is not the focus here).
  • Distinction Between Core and Default is Powerful: Representing concepts with both necessary (core) and typical (default) properties provides a richer and more nuanced knowledge representation.
  • Default Subsumption Bridges Theory and Practice: The novel definition of default subsumption is key to enabling practical default reasoning within a TL framework.
  • Interactive Diagnosis is Realistic: The paper highlights that real-world diagnosis is rarely a one-shot process. An interactive system that intelligently seeks further information is more practical and effective.

Key Results

  • Comparable Performance to Doctors: In an evaluation using 63 patient protocols, the DTL-based diagnosis system performed comparably to medical doctors (general practitioners and specialists) in cases where the medical community itself could agree on a diagnosis.
  • Successful Handling of Ambiguity: The system performed well even in cases where doctors required discussion to reach a consensus, suggesting its ability to handle less clear-cut scenarios.
  • Efficient Reasoning: The diagnostic engine's algorithm led to quick focusing on the right direction within the taxonomy, only wandering when no diagnosis was found.

Limitations and Future Work

  • Limited Expressivity: While DTL is more expressive than many TLs, it still has limitations compared to some theoretical frameworks. The paper notes that the focus is on defaults, not features with predicate values.
  • Approximate Default Classification: The current default classification algorithm is an approximation. The authors suggest it needs to be developed further for a more complete algorithm, especially for handling potential inconsistencies and reaching a fix-point.
  • Primitive vs. Defined Concepts: The paper questions the universal appropriateness of the strict distinction between primitive and defined concepts in TLs for application domains. It observes that in their application, most inheritance relationships were between core concepts, while properties were defined for default concepts, suggesting this might be a common pattern worth investigating.
  • Interactive Information Gathering: While the system uses heuristics for asking questions, a more sophisticated approach using information theory for optimal question selection is mentioned as a direction for future work.

Summary updated

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