The Effect of Preferences in Abstract Argumentation under a Claim-Centric View

Michael Bernreiter, Wolfgang Dvorák, Anna Rapberger, Stefan WoltranAAAI 2023 · DOI

TL;DR

This paper investigates how incorporating preferences into abstract argumentation frameworks affects the evaluation of claims, particularly when arguments with the same claim are treated uniformly. They introduce four ways to handle preferences and analyze their impact on semantic properties and computational complexity, finding that one method (Reduction 3) preserves desirable properties while others do not.

What Problem Does This Paper Address?

Formal argumentation systems are used to model reasoning and decision-making. A key aspect is how to handle arguments that might have varying degrees of plausibility or strength – this is where "preferences" come in. Traditionally, these systems evaluate arguments themselves. However, often we are more interested in the claims or conclusions that arguments support. Recent work has shown that shifting the focus from arguments to claims can change how we evaluate acceptability, especially if arguments with the same claim are treated identically. This paper tackles the question: what happens when we combine this claim-centric view with preferences? Specifically, they examine how different ways of incorporating preferences into argumentation frameworks affect the resulting claim-level evaluations, and whether the benefits of claim-centric views (like preserving certain desirable properties and having lower computational complexity) are maintained.

Methods

The authors build upon Claim-Augmented Argumentation Frameworks (CAFs), which associate claims with arguments. They focus on "well-formed" CAFs (wfCAFs), where arguments with the same claim have identical attack patterns, as these possess desirable properties. The core of their method involves applying four established "preference reductions" to these wfCAFs. These reductions modify the attack relation of the framework based on an asymmetric preference relation between arguments. For instance, one reduction might remove attacks from a weaker argument to a stronger one. They then analyze the resulting CAF classes for properties like I-maximality (ensuring extensions are not subsets of each other) and computational complexity of reasoning tasks (like checking if a claim is accepted).

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Preferences Break Well-Formedness: A crucial finding is that applying preference reductions to well-formed CAFs often results in frameworks that are no longer well-formed. This means the desirable properties of wfCAFs are not automatically preserved when preferences are introduced.
  • Reduction 3 is a "Sweet Spot": Among the four preference reduction methods studied, Reduction 3 (which removes attacks that contradict preferences only when the weaker argument is attacked by the stronger one) stands out. It preserves I-maximality for most semantics and maintains the computational complexity advantages of wfCAFs.
  • Trade-offs are Inevitable: For the other reductions (1, 2, and 4), there's a clear trade-off. While they might offer different expressiveness, they generally lead to a loss of I-maximality and, in some cases, increased computational complexity compared to wfCAFs.
  • Characterizing New CAF Classes: The paper introduces and characterizes four new classes of CAFs that arise from these preference reductions. These classes lie between the general CAF space and the more restricted wfCAF space, offering a nuanced understanding of how preferences shape argumentation structures.

Key Results

  • The paper defines four new classes of CAFs (R1-CAF, R2-CAF, R3-CAF, R4-CAF) resulting from applying preference reductions to wfCAFs. These classes are distinct and lie between wfCAFs and general CAFs.
  • I-Maximality: Reduction 3 preserves I-maximality for preferred, stable, semi-stable, and naive semantics (except for naive semantics, which is never I-maximal). Reductions 1, 2, and 4 generally fail to preserve I-maximality for these semantics.
  • Computational Complexity: For most semantics, reductions 2, 3, and 4 maintain the polynomial hierarchy complexity of verification problems seen in wfCAFs. However, Reduction 1 leads to the same complexity as general CAFs, losing the efficiency benefits. Complete semantics remain hard for reductions 2 and 4.

Limitations and Future Work

  • Transitive Preferences: While the paper considers asymmetric preferences, it also briefly analyzes the case of transitive preferences, noting that the relationships between the resulting CAF classes can change. A more in-depth study of transitive preferences could be beneficial.
  • Alternative Preference Handling: The paper focuses on "preference reductions" that modify the attack relation. Future work could explore other methods of incorporating preferences, such as lifting orderings over arguments to sets of arguments.
  • Alternative Semantics: The study could be extended to other argumentation semantics, particularly those that handle subset-maximization at the claim level.
  • Experimental Evaluation: An experimental evaluation of the theoretical results would be valuable.
  • More Structured Arguments: Incorporating more structure into arguments (beyond just claims) or exploring bipolar argumentation frameworks could be fruitful avenues.

Summary updated

BibTeX
@inproceedings{bernreiter2023effect,
  author = {Michael Bernreiter and Wolfgang Dvorák and Anna Rapberger and Stefan Woltran},
  title = {{The Effect of Preferences in Abstract Argumentation under a Claim-Centric View}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i5.25770},
}

Loading related content...

Figure