Dialectic reasoning with inconsistent information

Morten Elvang-Gøransson, Paul J. Krause, John FoxUAI 1993

TL;DR

This paper introduces a novel framework for "dialectic reasoning" that allows for the construction of arguments even in the presence of contradictory information within a knowledge base. By defining "acceptability classes" for arguments, the authors provide a way to quantify logical uncertainty and assign linguistic qualifiers, offering a more nuanced approach than traditional true/false logic.

What Problem Does This Paper Address?

In real-world scenarios, information is rarely perfectly consistent. Logical systems typically break down when faced with contradictions, leading to trivialization where anything can be proven. This paper addresses the challenge of how to reason effectively and draw meaningful conclusions from inconsistent knowledge bases, a common issue in artificial intelligence and human reasoning. The authors argue that humans are adept at localizing and managing inconsistencies, often suspending judgment on irrelevant contradictions. They aim to formalize this ability by developing a system that doesn't simply reject inconsistent information but rather uses it to modulate the "acceptability" of derived conclusions.

Methods

The core of the paper's contribution lies in its formalization of "dialectic reasoning" through a system of "argumentation." This system is built upon:

  • Labelled Deductive Systems (LDS): Formulas are augmented with labels that store information about their provenance (e.g., facts and warrants used to derive them). This is inspired by Toulmin's model of argumentation, where arguments carry information about their support.
  • Argumentation Consequence Relation: This defines how new arguments can be constructed from a database and a set of warrants. Arguments are represented as pairs of a conclusion and a set of labels of the facts they are based on.
  • Acceptability Classes: This is the key innovation. Instead of a binary true/false, arguments are assigned to one of several "acceptability classes" (A1 to A5). These classes are defined hierarchically, with stricter criteria for higher classes. Membership in these classes is determined by factors like the argument's existence, consistency, tautological nature, and importantly, the presence or absence of "defeating" arguments (rebutting or undercutting).
  • Linguistic Qualifiers: The acceptability classes are then mapped to linguistic terms (e.g., "supported," "plausible," "probable," "confirmed," "certain"). This allows for a qualitative representation of logical uncertainty, reflecting how confident we can be in a proposition.
  • Priorities: The paper also extends the framework to handle prioritized information, allowing explicit ordering of facts to resolve conflicts and refine acceptability.

Datasets

Not discussed in the paper. The paper uses illustrative examples of small, custom-defined databases (e.g., KM, KB1) to demonstrate the framework.

Key Insights and Ideas

  • Inconsistency as Uncertainty, Not Failure: The most significant insight is reframing inconsistency not as a system-breaking pathology, but as a source of logical uncertainty that can be quantified through argument acceptability.
  • Argument-Centric Reasoning: The paper shifts focus from propositions to arguments as the fundamental units of reasoning, acknowledging that the support for a conclusion is as important as the conclusion itself.
  • Graded Acceptability: The introduction of acceptability classes provides a spectrum of confidence, moving beyond binary truth values to better reflect nuanced reasoning.
  • Linguistic Qualifiers as a Bridge: Mapping acceptability classes to natural language terms makes the abstract logical concepts more accessible and interpretable, particularly for decision support.
  • Formalizing Human-like Reasoning: The framework aims to capture aspects of human reasoning, such as localizing inconsistencies and suspending judgment on irrelevant information.

Key Results

The paper presents a formal framework for dialectic reasoning with inconsistent information. While not focused on quantitative performance metrics, it demonstrates:

  • The ability to construct arguments for and against propositions from inconsistent databases without trivialization.
  • A hierarchy of acceptability classes that orders arguments based on logical criteria and defeat relationships.
  • A mapping from these acceptability classes to linguistic qualifiers, providing a qualitative measure of certainty.
  • An extension of the framework to incorporate explicit priorities between facts, refining the notion of acceptability.

The authors illustrate these concepts with examples showing how their system can assign different levels of "plausibility" or "confirmation" to conclusions based on the arguments that can be constructed. For instance, in one example, a conclusion is "confirmed" because no rebutting or undercutting arguments exist, even within an inconsistent database.

Limitations and Future Work

  • Qualitative Nature: The authors acknowledge that the assignment of linguistic qualifiers is an "interesting experiment" and not a definitive natural language semantics. They suggest that "implausible(p) might be better defined as plausible(-p)."
  • Complexity of Acceptability Criteria: While the paper defines several acceptability classes, the full implications and computational complexity of these definitions are not exhaustively explored.
  • Future Work: The authors suggest further exploration of the linguistic qualifiers, their integration with more sophisticated argumentation systems (especially those handling explicit priorities), and their application in decision support domains where numerical uncertainty coefficients are hard to elicit. They also note the potential for their framework to re-express existing formalisms like Poole's notion of specificity.

Summary updated

BibTeX
@inproceedings{elvanggransson1993dialectic,
  author = {Morten Elvang-Gøransson and Paul J. Krause and John Fox},
  title = {{Dialectic reasoning with inconsistent information}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1993},
}
Figure