Bilateral Gradual Semantics for Weighted Argumentation

Zongshun Wang, Yuping ShenAAAI 2024 · DOI

Abstract

Traditional abstract argumentation frameworks typically evaluate arguments along a single dimension of acceptability. This paper, presented at AAAI-24, introduces Bilateral Gradual Semantics, a framework that distinguishes between an argument's acceptability degree and its rejectability degree. By treating these as non-reciprocal values, the proposed model captures nuances in reasoning—such as the difference between a lack of support and active opposition—that are often lost in unilateral systems.

Introduction

In the field of Artificial Intelligence, formal logic and abstract argumentation provide the foundational tools for automated reasoning. While classical models categorize arguments as simply "accepted" or "rejected," gradual semantics introduced numerical scores to represent varying degrees of strength. However, these models are often "unilateral," assuming that rejectability is merely the inverse of acceptability. This paper argues that a bilateral perspective is necessary to model human-like common sense, particularly in complex domains like policy-making.

Problem Statement: The Need for Bilateral Evaluation

Existing unilateral models struggle to distinguish between arguments that lack evidence and those that face valid criticism. Consider two policies: one with moderate support and no opposition, and another with high support but a fatal flaw pointed out by critics. A unilateral system might assign them similar scores.

Figure 2: Simple weighted argumentation graphs showing different attack scenarios with initial weights.

As shown in Figure 2, arguments $a$ and $b$ face attacks from $x$ and $y$ respectively. In a bilateral framework, the final strength of $a$ and $b$ depends not just on their initial weights ($0.6$ and $0.8$), but on the distinct acceptability and rejectability of their attackers.

Methodology: Non-Reciprocal Degrees

The authors utilize Weighted Argumentation Graphs (WAGs), where arguments possess an intrinsic "basic weight." The core of the bilateral framework lies in the non-reciprocal relationship between two scores:

  1. Acceptability ($\sigma^+$): Calculated based on the argument's basic weight and both the acceptability and rejectability of its attackers.
  2. Rejectability ($\sigma^-$): Determined solely by the acceptability of its attackers.

This structure enables a "gradual defense" mechanism. If an attacker is itself highly rejectable due to a counter-attack, its impact on the original target is diminished.

Figure 4: A complex argumentation graph illustrating multi-level attacks and defense chains.

In Figure 4, the chain $d \to b \to a$ demonstrates this: argument $d$ attacks $b$, which in turn attacks $a$. The bilateral model evaluates how $d$'s attack on $b$ increases $b$'s rejectability, thereby protecting $a$'s acceptability through a "gradual defense."

Proposed Semantics and Results

The authors propose three specific semantics to handle the aggregation of attackers:

  • ARM (AR-max-based): Prioritizes the quality of the strongest attacker.
  • ARC (AR-card-based): Prioritizes the quantity of attackers.
  • ARH (AR-hybrid-based): Balances both quality and quantity through a compensation model.

Evaluating these against 23 logical axioms, the authors found that ARH is the most robust, satisfying nearly all principles. Furthermore, they proved that these bilateral scores are mathematically guaranteed to converge to a unique stable pair for any graph, including those with cycles.

Conclusion

By decoupling acceptability from rejectability, this research provides a finer-grained tool for decision-support systems, allowing AI to better reflect the complexities of human debate. Future work will explore extending this to bipolar graphs (including support relations) and collective attacks.

Summary updated

BibTeX
@inproceedings{wang2024bilateral,
  author = {Zongshun Wang and Yuping Shen},
  title = {{Bilateral Gradual Semantics for Weighted Argumentation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i9.28945},
}
Figure