Generalised Inference and Inferential Modelling

Raymond E. Jennings, Christine W. Chan, M. J. DowadIJCAI 1991

Abstract

Traditional logic is often constrained by the "conventional dogma" that its primary function is the preservation of truth. This paper challenges that notion, proposing a generalized framework where inference is understood as the preservation of various "metalinguistic properties" tailored to specific applications. By shifting the focus from universal truth to domain-specific "normality," the authors introduce "Inferential Structures" and "Models" that allow for disciplined reasoning even when dealing with inconsistent or uncertain information.

Introduction: Beyond Truth Preservation

In classical logic, the "material conditional" leads to the paradox of material implication: from a false premise, anything follows. This "explosive" nature makes classical systems brittle in real-world scenarios, such as AI expert systems or legal reasoning, where data sources are often conflicting. The authors argue for a move toward "principled" inference—reasoning that remains useful and disciplined even when inputs are not strictly "true" in a metaphysical sense.

Theoretical Framework: Inferential Structures and Models

The paper synthesizes three major non-classical logical traditions:

  1. The American Programme: Belnap’s four-valued logic designed for conflicting computer data.
  2. The Australian Programme: Paraconsistent logics that tolerate contradictions without collapsing.
  3. The Canadian Programme: A focus on preserving "levels of coherence" within a set of premises.

To unify these, the authors define Inferential Structures, which consist of a domain of objects and relations that preserve the "normality" of those objects' values. Building upon this, Inferential Models add a layer of "bespoke" logic, defining specific "normal ranges" for values and "preference orderings" that dictate which properties should be prioritized during reasoning.

Formal definition of the inferential relation within a model

Key Insights: Truth as a Designated Value

A central insight of this work is that "truth" in formal logic is often merely a label for a "designated value" (typically '1') that we wish to preserve. The authors argue that this label can be replaced with other properties like "level of coherence," "groundedness," or "safety."

Inference thus becomes a matter of bespoke design. For an industrial plant, "normal" might refer to safe operating temperatures; for a lawyer, it might mean "favorable to the client." The rules of inference are designed to preserve these specific definitions of normality across the reasoning process.

Formalization of inference as preservation across partitions of premises

Applications and Results

The authors demonstrate the utility of this framework through conceptual applications:

  • The Remoteness Advisor: A legal advisory system that identifies precedents by preserving "similarity relations" and "jurisdictional relations" in negligence cases.
  • Natural Inference: The suggestion that everyday reasoning (e.g., smoke implies fire) is a form of property preservation, specifically preserving the transitive closure of natural causal relations.

By framing disparate movements like modal logic and situation semantics under the "preservation paradigm," the paper provides a unified umbrella for non-classical reasoning.

Conclusion and Future Directions

The paper advocates for a shift in the role of the logician: from seeking a single "True" logic to observing how practitioners in fields like law and science actually reason, and then building formal systems that mirror those practices. While the framework is currently schematic and theoretical, it lays the groundwork for more robust, domain-specific automated reasoning systems. Future work must address the computational complexity of maintaining these complex preference orderings and the inherent subjectivity in defining what counts as "normal."

Summary updated

BibTeX
@inproceedings{jennings1991generalised,
  author = {Raymond E. Jennings and Christine W. Chan and M. J. Dowad},
  title = {{Generalised Inference and Inferential Modelling}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1991},
}
Figure