An Axiomatic Account of Formal Argumentation

Martin Caminada, Leila AmgoudAAAI 2005

Abstract

Formal argumentation systems provide a structured way for AI agents to reason through conflicting information. However, instantiating abstract frameworks often leads to logical anomalies where systems reach contradictory or incomplete conclusions. This paper proposes "rationality postulates"—specifically closeness and consistency—to serve as formal sanity checks. By employing closure operators and restricted rebutting, the authors provide a mathematical foundation for building logically sound argumentation engines that are guaranteed to remain rational across various semantics.

Introduction

Argumentation theory is a cornerstone of symbolic Artificial Intelligence, providing a framework for agents to navigate conflicting data by constructing arguments, identifying "defeats" between them, and selecting a set of justified conclusions. While Phan Minh Dung’s seminal 1995 work established an elegant abstract framework for this process, the practical implementation of these theories often reveals significant gaps. When abstract theories are instantiated with concrete logical rules, researchers frequently encounter "anomalies"—results that violate basic logical intuition.

For instance, a system might conclude that an individual is a bachelor while simultaneously failing to conclude that they are unmarried, despite having a strict rule defining the relationship between the two. Even worse, a system might conclude that an individual is both married and a bachelor. This paper addresses these failures by introducing an axiomatic approach to formal argumentation, establishing a set of "rationality postulates" similar to the AGM postulates used in belief revision.

The Rationality Postulates

The authors argue that any well-behaved argumentation system must satisfy two primary postulates to be considered rational:

  1. Closeness: If a system justifies a specific set of conclusions, that set must be closed under the system's own strict (non-exception) rules. If the system concludes $A$ and has a rule $A \to B$, it must also conclude $B$.
  2. Consistency: A system should never justify a set of conclusions that are mutually exclusive or contradictory (e.g., concluding both $p$ and $\neg p$).

Technical Methodology

To ensure these postulates are met, the authors explore two technical mechanisms: closure operators and the refinement of the defeat relation.

Closure Operators

The paper examines how strict rules within a knowledge base should be treated. They compare two primary methods:

  • Propositional Closure ($Cl_{pp}$): This treats rules as material implications. While it guarantees consistency, it often generates an excessive number of "garbage" rules that can clutter the argument space.
  • Transposition Closure ($Cl_{tp}$): This is a more surgical approach where a rule like $a \to b$ implies its transposition $\neg b \to \neg a$. The authors identify this as a "Goldilocks" solution that ensures logical conflicts are visible without over-complicating the system.

Restricted Rebutting

In many standard argumentation systems, a weak argument can "block" a strong one simply by contradicting its conclusion. The authors introduce Restricted Rebutting, which redefines the "defeat" relation. Under this approach, strict arguments are given precedence over defeasible (exception-prone) ones in specific configurations. This prevents logically inferior arguments from interfering with the acceptance of sound conclusions.

Key Insights and Results

A significant contribution of the paper is the identification of the "Self-Defeat" trap. The authors demonstrate that simply adding more logical rules (closure) can inadvertently create arguments that prevent themselves or other valid arguments from being accepted. This highlights the delicate balance required when designing rule-based systems.

Through a series of formal proofs, the authors establish the following:

  • Theorem 1: Propositional Closure ($Cl_{pp}$) guarantees closeness and consistency, but only under grounded semantics.
  • Theorems 2-4: The combination of Transposition Closure and Restricted Rebutting is more robust. It guarantees that every complete extension—including preferred, stable, and grounded semantics—satisfies both closeness and consistency.

These results provide a definitive "recipe" for developers: by applying transposition to strict rules and restricting the rebutting relation, they can mathematically guarantee the rationality of their argumentation engine.

Limitations and Future Work

While this axiomatic account provides a rigorous foundation, it currently assumes that all rules carry equal weight. In real-world applications, arguments often have different priorities (e.g., legal statutes vs. common sense). Future research is needed to integrate preference-based reasoning into these postulates. Additionally, the authors suggest exploring further postulates, such as those governing the "reinstatement" of arguments, and extending the framework to more complex first-order logics.

Summary updated

BibTeX
@inproceedings{caminada2005axiomatic,
  author = {Martin Caminada and Leila Amgoud},
  title = {{An Axiomatic Account of Formal Argumentation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2005},
}

Loading related content...

Figure