Ranking-based Argumentation Semantics Applied to Logical Argumentation

Jesse Heyninck, Badran Raddaoui, Christian StraßerIJCAI 2023 · DOI

TL;DR

This paper bridges the gap between abstract argumentation semantics and structured logical argumentation. It shows that applying ranking-based semantics to structured argument construction methods results in "culpability measures," which quantify the degree of responsibility of premises for inconsistencies. This offers a more nuanced understanding of argument acceptability than simple acceptance/rejection.

What Problem Does This Paper Address?

Formal argumentation systems aim to model reasoning with conflicting information. A key distinction exists between extension-based semantics, which determine sets of acceptable arguments, and ranking-based semantics, which assign a degree of acceptability to each argument. While ranking-based semantics have been extensively studied in the abstract setting (where arguments and their attacks are given), their application to structured argumentation (where arguments are built from logical formulas) has been largely unexplored. Structured argumentation is crucial because it grounds arguments in logical reasoning. This paper addresses the question of what logical meaning ranking-based semantics acquire when applied to structured argumentation frameworks, particularly those built from logical languages. It investigates how these semantics behave and whether they offer a useful way to understand the "culpability" or blame associated with premises in inconsistent arguments.

Methods

The authors systematically investigate the behavior of ranking-based semantics when applied to structured argumentation frameworks, specifically focusing on logic-based argumentation. They define and analyze simple contrapositive assumption-based frameworks (ABFs), which provide a general way to construct argumentation frameworks from strict and defeasible premises using a Tarskian logic.

Key methodological contributions include:

  • Generalizing Postulates: They adapt and generalize existing postulates for ranking-based semantics (like void precedence, monotony, and counter-transitivity) and culpability measures (like freeness, dominance, blame, and consistency) to the structured setting of ABFs.
  • Connecting Semantics to Culpability: A central finding is that a wide class of ranking-based semantics, when applied to structured argumentation frameworks, induce culpability measures. This means the ranking scores can be interpreted as a measure of how "responsible" an argument (or its premises) is for an inconsistency.
  • Analyzing Attack Rules: They examine how different attack rules (Defeat, Direct Defeat, Undercut, etc.) in structured argumentation influence the resulting rankings and culpability measures.

Datasets

Not discussed in the paper. The paper focuses on theoretical frameworks and postulates rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Ranking as Culpability: The most significant insight is that ranking-based semantics applied to structured argumentation frameworks provide a logical interpretation as culpability measures. This offers a more granular view of inconsistency than traditional methods, allowing us to pinpoint the "blame" associated with specific premises.
  • Robustness of Ranking Semantics: The paper demonstrates that ranking-based semantics are surprisingly robust to various choices in structured argumentation, such as different attack rules or the presence of strict premises. This suggests their broad applicability.
  • Falsity Postulate: A key finding is that many ranking-based semantics satisfy the "falsity" postulate, meaning they correctly rank self-contradictory arguments lower than others. This is a desirable property that some existing culpability measures lack.
  • Bridging Abstract and Structured: The work effectively bridges the gap between abstract argumentation theory and concrete logical reasoning, showing how theoretical semantics can gain practical meaning in structured settings.

Key Results

  • A large family of ranking-based semantics, when applied to structured argumentation frameworks derived from ABFs, induce culpability measures.
  • These induced culpability measures are shown to be more intuitive and robust than some existing culpability measures, particularly in handling self-contradictory premises. For instance, they can distinguish between a self-contradictory premise and a consistent one, which some other measures fail to do.
  • The paper shows that ranking-based semantics satisfying monotony, void precedence, and void best rank also satisfy the key postulates for culpability measures (Freeness, Dominance, Blame, Consistency).
  • The "falsity" postulate, which ensures that self-contradictory arguments are ranked lower, is satisfied by ranking-based semantics that also satisfy monotony.

Limitations and Future Work

  • Focus on Specific Logics: While the framework is general, the primary focus is on logics with specific properties (explosive, contrapositive, uniform). Extending to a wider range of logics might reveal different behaviors.
  • Comparison with Advanced Culpability Measures: The authors mention that future work will involve comparing their induced culpability measures with more advanced ones, such as tacit culpability measures.
  • Ranking Semantics over Sets of Arguments: The paper suggests extending the work to ranking semantics over sets of attacking logical arguments.
  • Intrinsic Strength: The authors note that their work can be extended to account for the "intrinsic strength" of arguments, a concept often assumed in ranking-based logics.
  • Infinite Argument Sets: For systems with support-conclusion pairs, infinite argument sets can arise. While the paper acknowledges this and mentions a strategy to handle it (using well-ordering), it's a complexity that might require further exploration.

Summary updated

BibTeX
@inproceedings{heyninck2023ranking,
  author = {Jesse Heyninck and Badran Raddaoui and Christian Straßer},
  title = {{Ranking-based Argumentation Semantics Applied to Logical Argumentation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/364},
}
Figure