A Comparative Study of Ranking-Based Semantics for Abstract Argumentation

Elise Bonzon, Jérôme Delobelle, Sébastien Konieczny, Nicolas MaudetAAAI 2016 · DOI

Abstract

This paper provides the first systematic comparison of ranking-based semantics in abstract argumentation. By evaluating six major frameworks against a suite of 15 logical properties, the authors reveal fundamental differences in how these systems value the "strength" of an argument, ultimately categorizing them into two distinct families based on how they treat defense branches.

Introduction

In classical abstract argumentation, as pioneered by Dung (1995), arguments are evaluated using a binary approach: they are either "accepted" or "rejected." While mathematically elegant, this "all-or-nothing" paradigm is often insufficient for complex real-world scenarios, such as large-scale web debates or legal reasoning, where a more nuanced understanding of argument strength is required. To address this, researchers have developed "ranking-based semantics" that produce a preference order of arguments. However, these methods were largely developed in isolation, leaving a gap in our understanding of their relative behaviors and the logical principles they satisfy.

Methodology: An Axiomatic Approach

The authors conduct a formal axiomatic study to compare six representative ranking-based semantics: Categoriser, Social Abstract Argumentation (SAF), Discussion-based, Burden-based, Tuples*, and Matt & Toni (M&T). Rather than proposing a new algorithm, the study formalizes 15 logical properties (axioms) from existing literature, such as:

  • Void Precedence: Non-attacked arguments should be ranked highest.
  • Cardinality Precedence: An argument with more attackers is weaker than one with fewer.
  • Quality Precedence: The strength of an argument depends on the strength of its attackers.

To illustrate the divergent rankings produced by these methods, the authors utilize a specific 5-argument graph as a running example.

Figure 1: The running example argumentation framework used to demonstrate divergent rankings across different semantics.

Key Insights: The Local vs. Global Split

The study identifies a fundamental divide in how semantics calculate argument strength. Some frameworks, like Categoriser and SAF, focus on the "local" quality or quantity of direct attackers. In contrast, others like Tuples* and M&T adopt a "global" view, considering the entire branch of attacks and defenses stretching back through the argumentation framework.

A critical discriminator identified is the Attack vs Full Defense (AvsFD) property. This property evaluates whether an argument attacked once by a "clean" (non-attacked) argument is worse than an argument defended by many branches.

Figure 2: Synthetic structures used to test properties like Cardinality Precedence and Attack vs Full Defense.

The research proves that certain desirable properties are mathematically incompatible. For instance, a semantics cannot simultaneously satisfy both Cardinality Precedence (counting attackers) and Quality Precedence (valuing attacker strength).

Results and Comparative Analysis

The primary output of this study is a comprehensive mapping of semantics to satisfied properties. Key findings include:

  • Baseline Properties: Six properties (Abstraction, Independence, Void Precedence, Non-attacked Equivalence, Totality, and Addition of Attack Branch) are satisfied by nearly all semantics and should be considered the standard for future systems.
  • Unique Frameworks: Tuples is highly unique but fails the Totality property (cannot compare all pairs) and struggles with cycles. Matt & Toni (M&T) is the only semantics to satisfy Self-Contradiction*.
  • Grounded Semantics: The classical binary approach is shown to be a "degenerate" version of ranking, lacking the nuance to rank arguments within the accepted or rejected sets.

Conclusion and Future Work

This systematic comparison clarifies the landscape of ranking-based semantics, identifying two main families based on their treatment of defense. Future work remains in refining semantics like Tuples* to handle cyclical graphs and identifying the specific real-world contexts (e.g., legal vs. social media) that justify the use of one set of axioms over another.

Summary updated

BibTeX
@inproceedings{bonzon2016comparative,
  author = {Elise Bonzon and Jérôme Delobelle and Sébastien Konieczny and Nicolas Maudet},
  title = {{A Comparative Study of Ranking-Based Semantics for Abstract Argumentation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2016},
  doi = {10.1609/aaai.v30i1.10116},
}
Figure