Introduction
In the late 1980s, Artificial Intelligence faced a significant hurdle known as the "brittleness" problem. Expert systems were highly effective within their narrow, pre-defined domains but failed completely when encountering information at the boundaries of their knowledge. To address this, researchers sought to develop systems capable of "graceful degradation" through plausible inference—the ability to draw conclusions that are likely true, even if not strictly guaranteed by formal logic.
Traditionally, semantic networks relied heavily on ISA (is-a) links for property inheritance (e.g., "A canary is a bird; birds have wings; therefore, a canary has wings"). However, human reasoning utilizes a far richer set of relations, such as CAUSES, CONTAINS, or SETTING-OF. This paper explores how systems can automatically derive these non-ISA rules and distinguish between plausible and implausible conclusions.
Methodology: Rule Derivation and Deep Structure
The authors proposed a two-step approach to move beyond simple inheritance. First, they treated any pair of relations ($R_1, R_2$) as a potential template for a rule. If node $n_1$ relates to $n_2$ via $R_1$, and $n_2$ relates to $n_3$ via $R_2$, the system attempts to infer a relation between $n_1$ and $n_3$.
To predict which of these rules would be perceived as "plausible" by humans, the researchers mapped 18 "surface" relations onto two "deep" semantic primitives:
- $h$ (Hierarchical): Relations involving part-whole structures or inclusion.
- $t$ (Temporal): Relations involving a sequence in time.
By analyzing these primitives, they identified two key structural properties: Transitivity (where the deep structure implies a consistent ordering) and Generalized Property Inheritance (GPI) (where a part inherits properties from its whole).
Key Insights: Structure over Content
A primary finding of the study is that the structure of a rule—its deep mapping—is often a better predictor of plausibility than the specific concepts involved. For instance, a "contains" relation followed by a "causes" relation often yields a plausible conclusion regardless of the specific items (e.g., alcohol in cough syrup).
Furthermore, for relations that are semantically ambiguous (like MECHANISM-OF, which can be viewed as either hierarchical or temporal), plausibility depends on maintaining a consistent interpretation across the entire inference chain. If a chain mixes a temporal interpretation with a hierarchical one, the resulting logic typically fails human plausibility tests.
Results: Plausibility and Chaining Decay
The study utilized the GRANT KB, a knowledge base of roughly 4,500 nodes, to generate over 3,000 specific inferences. These were then evaluated by human subjects. The results confirmed that transitive rules with consistent deep structures were significantly more plausible than intransitive ones, scoring nearly 50% higher on a combined metric.
However, the researchers also observed a "chaining decay" effect. "First-generation" inferences derived directly from the knowledge base had a mean plausibility score of 4.18/10. When these derived conclusions were used as premises for further "second-generation" inferences, the mean score dropped to 3.17/10. Despite this decay, the deep structure model remained highly predictive, matching human judgments 77% of the time for transitive items.
Conclusion and Limitations
While deep structure provides a powerful framework for plausible inference, it is not exhaustive. A significant portion of variance in human judgment stems from specific "item variance"—real-world nuances that abstract structures cannot capture. Additionally, subjectivity remains a challenge, as different individuals have different thresholds for plausibility based on their unique background knowledge. Future work will need to validate these structural principles across diverse knowledge bases to ensure their universal applicability.
Summary updated