Abstract
This paper introduces a computational framework for bridging the gap between rigid logical class definitions and the practical necessity of "best-fit" classification. By automatically transforming complex, rule-based criteria into "symbolically instantiated" prototypical instances known as KERNELs, the system enables efficient exact and inexact matching. This approach is particularly relevant for domains such as legal reasoning, where case facts rarely align perfectly with theoretical definitions.
Introduction
In classical AI, classification is typically a binary process: an instance either satisfies a set of logical criteria or it does not. However, in fields like law, medicine, or linguistics, a "closest match" is often required when an exact one is unavailable. Standard knowledge-representation languages often define classes using complex, pattern-directed inference rules. While robust, these rules are computationally expensive and difficult to use for measuring the "distance" between facts and definitions. This research proposes representing class criteria in the same format as the instances being classified, streamlining the comparison process.
Methodology: Automatic Transformation and KERNEL Generation
The core contribution is an algorithm for the automatic transformation of class definitions into a KERNEL structure. This process involves two primary components:
- Input Decomposition: Definitions are provided as a set of propositions (relations between typed variables) and pattern-directed inference rules that act as constraints.
- Symbolic Instantiation: The algorithm transforms these rules into a KERNEL—an "exemplary instance" containing only the propositions that must be true for class membership.
To manage the computational complexity of graph-based matching, the KERNEL is decomposed into small sub-networks. The system utilizes a staged matching function, DMATCH, which first tests these small components against case facts and then iteratively combines successful matches into larger structures.
Evaluation and Domain Application
The research focuses on the complex domain of legal and judicial reasoning, specifically utilizing the "TAXMAN" project. The primary test case involves the definition of BREORGANIZATION (corporate reorganization). The evaluation demonstrates the system's ability to match this intricate definition against a set of facts describing a legal case.
Key Insights
- Prototypical Representation: For matching purposes, a prototype (an instance) is often more effective than a raw rule set. Symbolically instantiating definitions allows for simpler matching algorithms that check for the presence of required components.
- Preprocessing Efficiency: The transformation from rules to a KERNEL is a one-time "compilation" process, significantly improving the efficiency of subsequent matching tasks.
- Handling Inexactness: The KERNEL provides a structural foundation for identifying "unknowns"—criteria that are neither explicitly satisfied nor denied. This allows the system to either re-interpret facts or adjust definitions to find the best fit.
Conclusion and Future Work
The paper successfully demonstrates that class definitions can be translated into a simpler KERNEL form for efficient exact matching. While the framework for inexact matching is established, the specific algorithms for computing "distance" from an exact match remain a subject for future implementation. Additionally, while the current work is tailored to legal reasoning, the KERNEL approach holds potential for broader applications in fields like speech and character recognition.
Summary updated