Evaluation of Arguments from Support Relations: Axioms and Semantics

Leila Amgoud, Jonathan Ben-NaimIJCAI 2016

Abstract

This paper introduces a comprehensive axiomatic framework for evaluating argument strength in graphs characterized solely by support relations. Moving beyond traditional attack-centric models, the authors propose three distinct semantics—Top-based (Tbs), Reward-based (Rbs), and Aggregation-based (Gbs)—which prioritize quality, quantity, or a balanced compensation of supporters, respectively. By defining 17 logical axioms, the study provides a formal foundation for aggregating an argument's intrinsic strength with the influence of its supporters, proving that different evaluative philosophies are mathematically incompatible.

Introduction

In classical AI argumentation theory, the focus is predominantly on negative relations where one argument defeats another. However, this approach often results in a "flatness" problem: in scenarios without attacks, all arguments are deemed equally strong, ignoring the nuances of positive reinforcement. Furthermore, existing models typically assume uniform intrinsic strength for all arguments. This research shifts the focus to "who supports whom," providing a formal method to calculate an argument's overall strength by aggregating its basic intrinsic value with the strengths of its supporters.

The Axiomatic Framework and Semantics

The authors employ an axiomatic approach, defining 17 properties that a robust evaluation method should satisfy. Key axioms include Monotony (more support increases strength), Strengthening (non-zero support must have an effect), and the distinction between Counting and Reinforcement. These axioms serve as a benchmark for three proposed semantics:

  1. Top-based Semantics (Tbs): Focuses on the "quality" of support, where strength is primarily determined by the single strongest supporter.
  2. Reward-based Semantics (Rbs): Prioritizes "quantity," where each additional supporter adds a diminishing reward.
  3. Aggregation-based Semantics (Gbs): Emphasizes "compensation," allowing many weak supporters to outweigh a single strong one.
A synthetic support graph (Example 1) used to demonstrate how different semantics rank argument strength based on topological relations.

Results and Comparative Analysis

The primary contribution is a rigorous comparison of these semantics against the 17 axioms. The analysis reveals fundamental trade-offs; for instance, Quality Precedence and Cardinality Precedence are mathematically incompatible. Tbs is the only model satisfying Quality Precedence, while Rbs favors Cardinality Precedence. Gbs serves as a middle ground, satisfying Compensation.

A summary of the axiomatic evaluation for the Tbs, Rbs, and Gbs semantics. Filled circles indicate satisfaction of the axiom, while open circles indicate failure.

All three semantics ensure stability by converging to a unique limit through iterative functions. This framework also establishes that support is not merely the inverse of an attack, requiring unique foundations such as the Non-Dilution axiom, which ensures that supporting another argument does not weaken the source argument itself.

Conclusion

This work provides a formal foundation for support-based argumentation, addressing the limitations of attack-only models. While currently restricted to support-only graphs, it sets the stage for future extensions into bipolar graphs where attacks and supports coexist. The framework's reliance on pre-defined "basic strengths" also opens avenues for research into objective initial weighting in practical applications like recommendation systems or social networks.

Summary updated

BibTeX
@inproceedings{amgoud2016evaluation,
  author = {Leila Amgoud and Jonathan Ben-Naim},
  title = {{Evaluation of Arguments from Support Relations: Axioms and Semantics}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure