Author: Wlodek W. Zadrozny (IBM Research, 1987)
Summary by: Research Blog Team
Abstract
In the late 1980s, the field of artificial intelligence struggled to reconcile formal logic with the fluid nature of human common sense. This paper critiques the then-dominant paradigm of circumscription—a method of non-monotonic reasoning based on minimizing "abnormality"—and proposes a novel architecture that incorporates a "referential level" of knowledge. By leveraging the structured information found in natural language dictionaries and grammars, Zadrozny demonstrates how AI systems can identify "intended models" of reality, effectively resolving long-standing logical paradoxes like the Yale Shooting Problem.
Introduction
In 1987, AI research was deeply invested in non-monotonic reasoning: the ability of a system to revise its conclusions when presented with new evidence. The classic example is the "Tweety" problem: if we know Tweety is a bird, we assume Tweety flies; if we later learn Tweety is a penguin, we must retract that assumption.
The standard mathematical tool for this was circumscription, introduced by John McCarthy. Circumscription works by telling a computer to assume that the world is as "normal" as possible—minimizing the number of entities that possess "abnormal" traits. However, as the field progressed, this purely logical approach hit a wall. It struggled with temporal paradoxes and required increasingly complex, manual "policy axioms" to function, leading researchers like Wlodek W. Zadrozny to seek a more intuitive, language-grounded alternative.
The Problem: Logic Without Context
Zadrozny identifies a fundamental flaw in traditional logic-based AI: it treats words as arbitrary symbols. To a standard logical engine, the predicates is_dead(x) and is_alive(x) are merely strings of characters. The system does not inherently understand that they are antonyms unless a programmer explicitly defines that relationship for every single problem.
This lack of inherent meaning led to significant failures, most notably the Yale Shooting Problem (Hanks and McDermott, 1986). In this thought experiment involving a sequence of actions (loading a gun, waiting, and shooting), standard circumscription often produced multiple, equally "logical" models. One model correctly predicted the victim would die, while another—equally valid to the computer—suggested the gun might spontaneously and mysteriously unload itself during the "wait" period to maintain the "normality" of the victim staying alive. Logic alone could not distinguish between a plausible reality and a ridiculous one.
The Proposed Solution: A Three-Level Architecture
Zadrozny proposes moving beyond simple minimization by introducing a three-tiered reasoning architecture:
- The Metalevel: This level defines the type of models the system prefers (e.g., the general principle of minimizing abnormality).
- The Object Level: This contains the specific facts of the current problem (e.g., "Tweety is a canary").
- The Referential Level: This is the paper’s primary innovation. It provides a situation-independent interpretation of symbols using "world knowledge" derived from natural language tools, such as the Longman Dictionary of Contemporary English and formal English grammars.
Instead of merely looking for the "smallest" logical model, the system searches for the intended model—the interpretation that best aligns with the linguistic constraints of the referential level.
Language as a Logical Constraint
The core insight of the paper is that natural language is not just a medium for communication, but a structured repository of common sense. By linking a reasoning engine to a dictionary, the system gains immediate access to a web of relationships. For example, if a dictionary defines a "canary" as a "bird," the referential level forces the logic to inherit bird-like properties (like flight) by default, without requiring a programmer to write specific abnormality axioms for every species.
Furthermore, Zadrozny argues that grammatical cues provide essential hints for logic. The word "unless," for instance, functions as a linguistic marker for the Closed World Assumption, signaling to the system exactly where it should apply non-monotonic reasoning.
Results and Applications
Zadrozny applies this framework to the "gold standard" problems of the era with notable success:
- Resolving the Yale Shooting Problem: By analyzing the grammatical structure of the problem description, the system naturally prefers the "intended" result (the victim dies). The linguistic context prevents the system from considering the "paradoxical" model where the gun unloads itself, as that interpretation violates the standard meaning of the actions described.
- Improved Inheritance: The system can deduce properties of specific objects (like Tweety) more efficiently by using the hierarchical structure of language, making the underlying logic much simpler and less prone to the "policy axiom" bloat that plagued earlier circumscriptive methods.
Limitations and Conclusion
While groundbreaking, the paper acknowledges several hurdles that remained in 1987. The technology to automatically parse an entire dictionary into formal logic was still in its infancy. Additionally, Zadrozny noted that while his framework offered a path forward for the "Frame Problem" (understanding what remains unchanged after an action), a complete solution remained an open question.
Ultimately, Zadrozny’s work prefigured a major shift in AI: the realization that true common sense cannot be engineered through pure logic alone. It requires a bridge to the vast, pre-existing structure of human language—a concept that continues to resonate in the era of modern large language models.
Summary updated