Parametrized Gradual Semantics Dealing with Varied Degrees of Compensation

Dragan Doder, Leila Amgoud, Srdjan VesicIJCAI 2023 · DOI

TL;DR

This paper introduces a novel, flexible framework for "gradual semantics" in argumentation systems, allowing for a spectrum of how weak attacks can compensate for strong ones. By introducing a parameter 'α', the authors unify existing semantics and create new ones that can handle compensation at various degrees, offering a more nuanced and comprehensive approach to evaluating argument strength.

What Problem Does This Paper Address?

In argumentation systems, we often need to determine the "strength" of an argument. This is usually done by analyzing its attackers. A key challenge arises when an argument is attacked by a few very strong attackers versus another argument attacked by many weaker ones. Existing approaches have struggled to precisely define how these different scenarios should be handled. Some principles, like "Cardinality Precedence" (CP) and "Quality Precedence" (QP), offer extreme strategies (favoring quantity or quality, respectively). A more nuanced approach, "Compensation," suggests that multiple weak attacks can balance out a single strong one. However, existing formalizations of compensation lack specificity, leaving it unclear how much compensation is occurring or when it's applicable. This paper aims to bridge this gap by providing a parameterized family of semantics that can handle compensation at a wide range of degrees, offering a more granular and unified perspective.

Methods

The core contribution is a novel, parameterized family of gradual semantics, denoted as $S_\alpha$. This family is defined by a parameter $\alpha \in (0, +\infty)$. The strength of an argument is calculated iteratively, starting with its basic weight and then updating based on the strengths of its attackers. The update function $F^\alpha$ incorporates the parameter $\alpha$, which controls the balance between the number of attackers and their individual strengths.

  • Parameterization: The parameter $\alpha$ acts as a dial:
  • As $\alpha \to 0$, $S_\alpha$ approaches a semantics that prioritizes the number of attackers (Cardinality Precedence).
  • As $\alpha \to +\infty$, $S_\alpha$ approaches a semantics that prioritizes the quality (strength) of attackers (Quality Precedence).
  • For any finite $\alpha > 0$, $S_\alpha$ exhibits compensation at varying degrees.
  • Unified Framework: The paper demonstrates that this parameterized family unifies several existing semantics, including those that satisfy compensation.
  • Formal Characterization: The authors provide a recursive definition for the semantics and prove its well-definedness and convergence. They also offer an alternative, direct characterization of the semantics using a system of equations.
  • Compensation Degrees: Three specific classes of compensation degrees are formally defined:
  • (n, i)-Compensation: A specific number of weak attackers compensating for a single strong attacker.
  • (m, c, n, d)-Compensation: A fixed number of attackers at a fixed strength compensating for another fixed number of attackers at a different fixed strength.
  • (m, $\bar{c}$, n, $\bar{d}$)-Compensation: A more general form where sets of attackers with varying strengths and numbers are compared.

Datasets

Not discussed in the paper. The paper uses example argumentation graphs (G1, G2, G3) to illustrate its concepts and results, but these are not presented as formal datasets.

Key Insights and Ideas

  • Granular Compensation: The most significant insight is the creation of a spectrum of compensation, rather than a binary "yes/no." The parameter $\alpha$ allows for fine-tuning how much compensation is applied, making the semantics more adaptable to different real-world scenarios.
  • Unification of Existing Work: The parameterized family elegantly unifies several previously distinct semantics. This provides a more coherent theoretical landscape for argumentation research.
  • Extreme Cases as Limits: It's insightful that the extreme values of the parameter $\alpha$ naturally lead to the well-established Cardinality and Quality Precedence principles, demonstrating the robustness of the framework.
  • Comprehensive Compensation Power: The paper argues that their novel family offers greater compensation capabilities than existing semantics, being able to satisfy all proposed compensation degrees, whereas others are limited.
  • Incompatibility of Semantics: The proof that different values of $\alpha$ lead to incompatible semantics highlights how sensitive argumentation evaluation can be to subtle differences in how compensation is handled.

Key Results

  • The proposed parameterized family $S_\alpha$ can handle compensation at various degrees, unifying existing semantics and offering a more comprehensive approach.
  • The semantics $S_\alpha$ satisfy Cardinality Precedence as $\alpha \to 0$ and Quality Precedence as $\alpha \to +\infty$.
  • The family $S_\alpha$ can satisfy all three defined classes of compensation degrees: (n, i)-Compensation, (m, c, n, d)-Compensation, and (m, $\bar{c}$, n, $\bar{d}$)-Compensation, for specific values of $\alpha$.
  • Existing semantics like Trust-based and Weighted h-Categoriser have limitations in their compensation capabilities compared to the proposed family. For instance, Trust-based semantics cannot satisfy any of the defined compensation degrees, and Weighted h-Categoriser is limited to specific conditions.

Limitations and Future Work

  • Parameter Selection: While the paper provides a framework with a parameter $\alpha$, it doesn't offer guidance on how to choose the "best" value of $\alpha$ for a given application. This remains an open question.
  • Computational Complexity: The paper focuses on the theoretical properties of the semantics. The computational complexity of calculating argument strengths for large graphs is not explicitly discussed.
  • Future Work: The authors suggest investigating other reasonable degrees of compensation and applying their formal approach to multiple criteria decision-making problems. They also mention exploring other principles for semantics beyond those discussed.

Summary updated

BibTeX
@inproceedings{doder2023parametrized,
  author = {Dragan Doder and Leila Amgoud and Srdjan Vesic},
  title = {{Parametrized Gradual Semantics Dealing with Varied Degrees of Compensation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/354},
}

Loading related content...

Figure