Osamu Imatcht and Yuji

IJCAI 1995

TL;DR

This paper bridges the gap between two major approaches to reasoning with inconsistent information in AI: coherence-based methods that revise knowledge to maintain consistency, and argumentation-based methods that manage conflicting arguments. The authors establish a formal correspondence, showing that coherence-based entailment can be recovered from argumentation systems, unifying these seemingly disparate lines of research.

What Problem Does This Paper Address?

Artificial intelligence systems often encounter conflicting pieces of information, leading to inconsistent knowledge bases. Reasoning with such bases is a significant challenge. Two primary paradigms have emerged to tackle this:

  1. Coherence-based approaches: These methods aim to restore consistency by selecting "preferred" subsets of the original knowledge base (e.g., maximal consistent subbases) and then applying classical logical entailment to these consistent subsets. The challenge here lies in defining what constitutes a "preferred" subset and how to select among them.
  2. Argumentation-based approaches: These methods accept inconsistency and focus on constructing arguments for and against a conclusion. Reasoning then involves determining the "acceptability" of these arguments, often by seeing if arguments supporting a conclusion can defeat arguments for its negation. The core difficulty is defining what constitutes a "defeat" relationship between arguments and how to determine overall acceptability.

While both approaches aim to handle inconsistency, they have developed largely independently. This paper seeks to unify them by demonstrating a deep connection between them.

Methods

The paper introduces and formalizes two argumentation systems based on different "defeat" relations between arguments:

  • "Rebuts" and "Undercuts": These are defined based on logical equivalence of conclusions or negation of support elements.
  • "Disagrees" and "Counterargues": These are weaker relations based on inconsistency of conclusions or sub-argument conflicts.

The authors then explore "acceptability classes" of arguments (e.g., arguments that are not rebutted, not undercut) and "stable extensions" of argumentation systems (maximal sets of mutually non-defeating arguments). The core methodological contribution is demonstrating that:

  1. Stable extensions of the "undercuts" argumentation system directly correspond to maximal consistent subbases of the original knowledge base. This is a key insight, linking the abstract notion of argument acceptability to the concrete task of selecting consistent knowledge subsets.
  2. Coherence-based inference relations (strong, confirmed, probable, plausible consequences) can be precisely restated in terms of these stable extensions. This establishes the formal correspondence.

Datasets

Not discussed in the paper. The paper focuses on formal methods and theoretical connections rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • A Formal Bridge: The paper provides a crucial formal link between coherence-based and argumentation-based reasoning, showing they are not fundamentally different but rather different perspectives on the same underlying problem of handling inconsistency.
  • "Undercuts" as the Key: The "undercuts" relation in argumentation systems is shown to be particularly powerful, directly mirroring the structure of maximal consistent subbases. This suggests that certain definitions of argument defeat are more fundamental for this connection.
  • Unifying Inference Relations: The paper demonstrates how established coherence-based inference relations (like "probable consequence") can be recovered and understood through the lens of argumentation system extensions. This offers a unified framework for understanding various non-monotonic inference mechanisms.
  • Acceptability Classes as Subbases: The acceptability classes of arguments (e.g., AR+(E), AR++(E)) are shown to have direct relationships with the properties of consistent subbases, providing an intuitive interpretation for these classes.

Key Results

  • Stable Extensions = Maximal Consistent Subbases: For the argumentation system using the "undercuts" relation, stable extensions are in one-to-one correspondence with maximal consistent subbases of the knowledge base.
  • Recovering Coherence-Based Entailment: The paper shows that plausible, probable, and confirmed consequences (from coherence-based approaches) can be precisely defined using the stable extensions of the "undercuts" argumentation system. For example, a formula is a "plausible consequence" if it is entailed by at least one stable extension.
  • Hierarchy of Conservatism: A table is presented illustrating a hierarchy of consequence relations, ordered from most conservative (e.g., "strong consequence") to least conservative (e.g., "plausible consequence"), showing how they relate to each other and to argumentation concepts.

Limitations and Future Work

  • "Counterargues" System: While the "undercuts" system yields a clean correspondence, the paper notes that the "counterargues" system does not have such a direct mapping to maximal consistent subbases, and stable extensions in this system can arise from inconsistent supports.
  • Preference Relations: The authors highlight that integrating preference relations (e.g., based on specificity or priority axioms) into argumentation systems is a significant area for future work. This could lead to more nuanced comparisons of arguments and potentially new inference relations.
  • Comparison of Arguments: Further research is needed on how to compare arguments supporting the same conclusion, especially when dealing with priority relations on the belief base.
  • New Acceptability Classes: Exploring acceptability classes based on preferences (e.g., arguments not "rebutted" by a preferred argument) could lead to new inference relations.
  • Practical Implementation: While the theoretical connections are strong, practical implementation and computational complexity of these unified approaches remain important considerations.

Summary updated

BibTeX
@inproceedings{1995osamu,
  title = {{Osamu Imatcht and Yuji}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1995},
}
Figure