Inheritance Comes of Age: Applying Nonmonotonic Techniques to Problems in Industry

Leora Morgenstern

Introduction

For decades, nonmonotonic reasoning (NMR)—the logic of "plausible" reasoning where conclusions can be retracted as new information emerges—has been a staple of artificial intelligence research. However, much of this work has remained confined to "toy problems," such as determining if a specific bird can fly. This paper argues that for NMR to remain relevant, it must address the messy, exception-heavy complexities of industrial applications. The author presents a practical bridge: Formula-Augmented Semantic Networks (FANs), applied to the intricate domain of medical insurance benefits.

The Industrial Gap: Beyond Toy Problems

In the late 1990s, while expert systems were finding success in niche industrial roles, NMR researchers were often preoccupied with abstract puzzles like the Yale Shooting Problem. Meanwhile, industries like medical insurance were struggling with massive knowledge bases of nested, often conflicting rules.

The primary barrier to industrial adoption was twofold: computational intractability (many NMR theories are NP-hard or undecidable) and a lack of robust tools. Morgenstern addresses this by focusing on a "non-toy" domain—medical insurance benefits inquiry—where the "80-20 rule" (getting it right 80% of the time) is insufficient. In insurance, a 20% error rate in benefit explanation is unacceptable, necessitating a system that can handle exceptions reliably.

A complex taxonomy of medical services illustrating the scale of industrial knowledge bases.

Formula-Augmented Semantic Networks (FANs)

To solve these real-world problems, the paper introduces the FAN, a hybrid knowledge structure that combines the efficiency of taxonomic inheritance with the expressive power of first-order logic.

1. Taxonomic Inheritance

A FAN begins with a standard inheritance network, typically a directed acyclic graph representing "is-a" relationships. This allows for a natural organization of services, such as Spinal Manipulation being a subtype of Physical Therapy.

2. Wff-Attachment

The core innovation of the FAN is wff-attachment. Instead of inheriting simple atomic attributes, nodes in the network can have well-formed formulae (wffs) of first-order logic attached to them. These formulae represent complex business rules, such as specific copay percentages or visit limits that apply to certain categories of service.

A FAN showing how specific rules (wffs) are attached to nodes in a therapy service hierarchy.

3. The Inheritance Algorithm

Determining which rules apply to a specific query involves an iterative upward traversal of the network. The system starts at the most specific node and moves toward the root, collecting wffs along the way. To maintain consistency, the algorithm favors the most specific information (the specificity principle) and computes preferred maximally consistent subsets of rules when contradictions arise.

Key Insights and Industrial Success

The research highlights several critical insights for applying AI in industry:

  • Structured Complexity: While commonsense reasoning is notoriously difficult, industrial rules are often highly structured. This structure makes them amenable to tractable NMR sub-theories.
  • Efficiency in Practice: Although inheriting complex logic is theoretically NP-hard, the actual sets of rules at any given node in an insurance database are small enough that the iterative algorithm performs efficiently.
  • User Acceptance: The resulting Benefits Inquiry System (BIS) was deployed for a large medical insurance corporation. It received excellent reviews from Customer Service Representatives (CSRs) and Policy Modifiers (PMs), demonstrating that NMR can handle "industrial-strength" complexity.

Conclusion and Future Directions

The success of the BIS proves that nonmonotonic techniques are ready for industrial application. However, challenges remain. The "knowledge engineering bottleneck"—the manual process of translating natural-language regulations into formal logic—is a significant hurdle. Furthermore, the industry lacks general-purpose tools for inheritance with exceptions, forcing developers to build custom engines. Future work will explore applying FANs to other domains like legal reasoning and medical treatment protocols, where structured exceptions are equally prevalent.

Summary updated

BibTeX
@inproceedings{morgenstern1997inheritance,
  author = {Leora Morgenstern},
  title = {{Inheritance Comes of Age: Applying Nonmonotonic Techniques to Problems in Industry}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1997},
}
Figure