Higher-Order Argumentation Frameworks: Principles and Gradual Semantics

Leila Amgoud, Dragan Doder, Marie-Christine Lagasquie-SchiexIJCAI 2024

Introduction

In classical abstract argumentation, arguments are represented as nodes in a graph, and attacks are directed edges between them. This "flat" structure assumes that an attack is an absolute, immutable fact. However, real-world reasoning is often recursive: we do not just attack arguments; we attack the validity of attacks themselves. This paper introduces the first "gradual semantics" for higher-order argumentation frameworks (HO-AFs), providing a formal, principled way to evaluate the strength of claims in complex, recursive reasoning scenarios.

The Challenge of Higher-Order Attacks

A higher-order attack occurs when an argument targets the connection between two other entities. For example, if Argument B attacks Argument A, a third Argument C might attack the relevance of B's criticism. While existing frameworks like RAF and AFRA handle these structures, they typically rely on binary "extension-based" semantics where an argument is either accepted or rejected. This paper addresses the lack of methods that assign a continuous degree of acceptance (e.g., a value from 0 to 1) to both arguments and attacks.

An example of a Higher-Order Argumentation Framework (HO-AF) where both arguments (nodes) and attacks (edges) are assigned numerical weights.

Foundational Principles and Semantics

The authors propose a comprehensive framework for evaluating HO-AFs through several novel contributions. They define nine foundational principles—including Equivalence, Maximality, Reinforcement, and Attack-sensitivity—that any gradual semantics should ideally follow. A key addition is the Equality principle, which examines whether attacks and arguments should be treated as mathematically identical entities.

To satisfy these principles, the authors introduce several families of semantics:

  • Structure-based semantics: These extend existing recursive frameworks by assigning values based on an element's membership in various acceptable sets.
  • Weighted Extensions (Hbs and Mbs): These adapt the H-Categorizer and Max-based semantics to higher-order settings using aggregation operators (such as average, min, or max) to combine the strength of an attack with the strength of its source.
  • Hybrid Semantics: A parameterized family that allows for "asymmetric" treatment, letting users decide if arguments should have precedence over attacks (A-Precedence) or vice versa (R-Precedence).

Comparative Analysis and Key Results

The paper provides the first formal comparison between major higher-order frameworks, revealing fundamental disagreements. For instance, AFRA may assign a value of 0 to an attack that is not even targeted by another argument, whereas RAF treats it differently.

The study culminates in a detailed "catalogue" of semantics, evaluating which models satisfy the proposed principles.

A comprehensive comparison of various semantics against the proposed foundational principles. The m-avg variant of Hbs emerges as a particularly robust choice.

As shown in the results, the m-avg (modified average) version of the H-Categorizer semantics emerges as a particularly robust choice, satisfying almost all proposed principles, including the challenging Attack-sensitivity and Reinforcement properties.

Conclusion and Future Work

This research formalizes the treatment of attacks as "first-class citizens" with their own degrees of seriousness. While the current model focuses on attacks, future work aims to incorporate "support" relations—where arguments or attacks strengthen one another—into the higher-order gradual framework. Additionally, the authors seek to fully characterize these semantics to prove that specific sets of principles uniquely define specific models.

Summary updated

BibTeX
@inproceedings{amgoud2024higher,
  author = {Leila Amgoud and Dragan Doder and Marie-Christine Lagasquie-Schiex},
  title = {{Higher-Order Argumentation Frameworks: Principles and Gradual Semantics}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure