Abstract
This paper provides a formal theoretical foundation for "gradual semantics" in argumentation—systems that calculate the strength of an argument based on its attackers. The authors introduce a novel framework that accounts for similarity and redundancy between attackers, ensuring that multiple redundant arguments do not unfairly weaken a claim more than a single unique one.
Introduction
In computational argumentation, gradual semantics are used to assign a numerical strength to arguments. Most existing models follow a principle called Counting: the more attackers an argument has, the weaker it becomes. However, a significant flaw exists: if attackers are nearly identical (e.g., three people using the same redundant evidence), standard systems treat them as independent blows. This paper addresses the challenge of mathematically adjusting attacker strength based on similarity before aggregation, ensuring rational and consistent evaluation.
Methodology: The Evaluation Method (EM)
The authors propose that any similarity-aware semantics can be decomposed into a three-step Evaluation Method (EM) consisting of three functions:
- Adjustment ($n$): This novel component modifies attacker strengths based on their mutual similarities to remove redundancy.
- Aggregation ($g$): This combines the adjusted strengths into a single "group" attack value.
- Influence ($f$): This calculates the final strength of the target argument by weighing its initial value against the total impact of its attackers.
The paper defines "well-behaved" functions through intuitive constraints (e.g., zero-strength attackers have no impact). They introduce the Parameterized Function ($n^\rho_{max}$), which ranks attackers and discounts each based on its maximum similarity to those already considered.
Illustrative Examples and Sanity Checks
As a theoretical study in formal logic, this work utilizes illustrative argumentation graphs to demonstrate that the proposed semantics align with human-like reasoning regarding redundancy.
These graphs allow for "sanity checks" of the model. For instance, the model ensures that an attacker completely dissimilar to others retains its full initial value (Independence Condition) and that the framework generalizes the classic h-Categorizer semantics when all arguments are 100% dissimilar.
Key Results and Theoretical Contributions
The authors successfully defined a broad family of semantics satisfying key rationality principles:
- Reinforcement: Stronger attackers result in a weaker target.
- Monotony: More dissimilar attackers result in a weaker target.
- Sensitivity to Similarity: Redundant information is appropriately discounted.
Mathematically, the paper proves that for any argumentation framework, a unique solution (a set of argument strengths) exists under "well-behaved" conditions. Theorems 3-7 establish the necessary mathematical constraints on the $f, g, \text{ and } n$ functions to guarantee these logical properties.
Conclusion and Future Work
While the $n^\rho_{max}$ function is a significant step, the authors acknowledge its "coarse" nature as it only considers the most similar neighbor. Future research will explore "fair" adjustment functions accounting for cumulative similarity. Additionally, the authors plan to apply these semantics to practical domains such as analogical reasoning and online debate platforms, where redundancy is common.
Summary updated