On the Graded Acceptability of Arguments

Davide Grossi, Sanjay ModgilIJCAI 2015

Introduction

In symbolic Artificial Intelligence, Abstract Argumentation Frameworks (AFs)—pioneered by Phan Minh Dung in 1995—serve as the foundational model for conflict resolution. In a standard Dung graph, arguments are represented as nodes and attacks as directed edges. Traditionally, this system is binary: an argument is either "justified" or "rejected" based on whether its attackers are themselves attacked (defended).

However, this binary approach fails to capture the intuitive notion that the quantity of evidence matters. An argument attacked by ten independent sources is qualitatively "weaker" than one attacked by only one, even if both are eventually defended. Conversely, an argument supported by multiple independent defenders feels more robust. This paper addresses the challenge of formalizing "strength" using only the graph's inherent structure, avoiding the need for external numerical weights or subjective probabilities.

Methods: Graded Defense Functions

The authors build their theory on two core principles: more attackers make an argument weaker, and more defenders make an argument stronger. To implement this, they generalize the standard defense function into a Graded Defense Function ($d_n^m$).

In this framework, two integers serve as thresholds:

  • $m$ represents the threshold for attackers.
  • $n$ represents the threshold for defenders.

An argument is considered $m,n$-defended if it has fewer than $m$ attackers that are not themselves counter-attacked by at least $n$ arguments. By varying these integers, the authors create a "ladder" of justification. Using fixpoint theory (specifically the Knaster-Tarski theorem), they demonstrate that one can iteratively apply these graded functions to reach stable sets of arguments, mirroring Dung’s original theory but with finer granularity.

The space of graded defense defined by thresholds m and n.

Key Insights and Structural Ranking

A significant contribution of this work is the discovery that the topology of a conflict graph contains an inherent hierarchy. One does not need to assign arbitrary probabilities to determine strength; the "accrual" of defenders provides a structural basis for ranking.

One critical mathematical constraint identified is the $n \ge m$ boundary. For graded semantics to remain "well-behaved" and follow the fundamental lemmas of argumentation theory, the number of defenders ($n$) must generally be greater than or equal to the number of attackers ($m$). If attackers outnumber defenders, the system can become unstable or fail to produce valid extensions.

Key Results

The authors proved that their graded framework is a true generalization of Dung’s; setting $m=1$ and $n=1$ recovers the classic 1995 results. Furthermore, the paper demonstrates how to produce a partial ordering of arguments through Hasse diagrams.

A Hasse diagram (vi) showing the partial ordering of argument sets based on graded acceptability.

As shown in the figure above, the framework can formally prove that certain arguments are "more justified" than others based on their position in the graph, even if they would all be labeled simply "justified" in a binary system. This successfully models "accrual"—the idea that multiple independent reasons for a conclusion make it stronger—within an abstract framework, a feat previously thought to require more complex, structured logic.

Limitations and Future Work

Because the system relies on a partial order, some arguments remain "incomparable." Future research may explore lexicographic ordering to force a total ranking, though this may impact logical purity. Additionally, while the mathematical foundations are robust, behavioral experiments are needed to determine if this counting logic aligns with human perceptions of argument strength in natural debate.

Summary updated

BibTeX
@inproceedings{grossi2015graded,
  author = {Davide Grossi and Sanjay Modgil},
  title = {{On the Graded Acceptability of Arguments}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure