Revisiting Unrestricted Rebut and Preferences in Structured Argumentation

Jesse Heyninck, Christian StraßerIJCAI 2017 · DOI

TL;DR

This paper tackles a fundamental challenge in structured argumentation: how to allow arguments to "rebut" (attack) each other's conclusions freely, even if those conclusions are derived from strict logical rules, without breaking essential properties like consistency and non-interference. The authors introduce a novel "generalized rebut" mechanism that allows arguments to attack if their conclusion claims a subset of the attacked argument's sub-conclusions are untenable together, thus resolving a key dilemma in argumentation theory.

What Problem Does This Paper Address?

Structured argumentation frameworks are powerful tools for modeling defeasible reasoning – reasoning that can be revised in light of new information. A key aspect of these frameworks is how arguments can attack each other. One type of attack is "rebut," where an argument attacks another by concluding the opposite of its conclusion.

The standard ASPIC+ framework restricts rebuts: an argument cannot be rebutted if its conclusion is derived using a strict rule (like those from classical logic). While this avoids certain paradoxes, it's considered counter-intuitive in some "dialectical" contexts where free exchange of arguments is natural. An alternative, ASPIC⁻, allows "unrestricted rebut," meaning any argument can rebut any other. However, ASPIC⁻ suffers from "contamination problems," violating crucial rationality postulates like consistency and non-interference, especially when strict rules are domain-independent (e.g., based on classical logic). This means adding irrelevant information can lead to contradictory conclusions or change valid conclusions.

The paper addresses the dilemma: how can we achieve the intuitive appeal of unrestricted rebut without sacrificing essential logical properties, particularly when dealing with domain-independent strict rules?

Methods

The core contribution is the introduction of a new attack relation called generalized rebut (GeRe). This generalizes the concept of unrestricted rebut.

Here's how it works:

  • Standard Unrestricted Rebut (UrRe): Argument a unrestrictedly rebuts b if conc(a) = ¬conc(b) and b is defeasible.
  • Standard Restricted Rebut (ReRe): Argument a restrictedly rebuts b if conc(a) = ¬conc(b) and b's conclusion was derived using a defeasible rule.
  • Generalized Rebut (GeRe): Argument a generalized rebuts b if b is defeasible and conc(a) expresses that a subset of the conclusions of b's subarguments cannot all hold simultaneously. This is formalized by defining C(b) as the set of conclusions of b's subarguments, and a GeRe b if conc(a) is equivalent to ¬(c1 ∧ c2 ∧ ... ∧ cn) where {c1, ..., cn} is a subset of C(b).

The authors also introduce a mechanism to handle priorities between defeasible rules using total pre-orders. They then demonstrate that their proposed system, ASPIC, with generalized rebut and these priority mechanisms, satisfies the four rationality postulates (Direct Consistency, Closure, Indirect Consistency, and Non-Interference) under certain conditions on the strict rules (closure under Transposition, Resolution, and Cut).

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • The "Contamination" Problem: The paper highlights a critical issue in argumentation systems: how to prevent irrelevant information from corrupting reasoning, especially when strict rules are very general (like classical logic). Unrestricted rebut, while intuitive, exacerbates this.
  • Generalized Rebut as a Solution: The core insight is that the problem with unrestricted rebut isn't just about direct negation, but about arguments that express the incompatibility of multiple conclusions. Generalized rebut captures this by allowing an argument to attack another if its conclusion states that a set of the attacked argument's sub-conclusions cannot all be true together.
  • Resolving the Dilemma: The paper shows that by generalizing rebut, they can avoid the counterexamples that plagued previous attempts to achieve unrestricted rebut while maintaining desirable logical properties. This is a significant step towards more robust and intuitive argumentation systems.
  • Importance of Strict Rule Properties: The paper emphasizes that the desirable properties (like non-interference) are achieved when the underlying strict rules have certain logical properties (Transposition, Resolution, Cut). This suggests that the interaction between strict and defeasible reasoning is crucial.

Key Results

The main result is the definition of the ASPIC system with generalized rebut, which is proven to satisfy the four rationality postulates (Direct Consistency, Closure, Indirect Consistency, and Non-Interference) for argumentation systems where the strict rules are closed under Transposition, Resolution, and Cut, and priorities over defeasible rules form a total preorder. This is presented as a significant theoretical advance for structured argumentation frameworks.

Limitations and Future Work

  • Restrictions on Strict Rules: The proofs of the rationality postulates rely on the strict rule base being closed under Transposition, Resolution, and Cut. The authors acknowledge that exploring what happens when these conditions are relaxed is a direction for future work.
  • Total Pre-orders: The paper assumes total pre-orders for priorities. Investigating non-total orders is also suggested.
  • Other Lifting Principles: The paper mentions exploring other "lifting principles" (ways to extend preferences from defeasible rules to arguments) like "last link."
  • Variants of Generalized Rebut: The authors note that generalized rebut, as defined, might still lead to counter-intuitive results in some specific cases (e.g., when an attacker is weaker than the attacked argument's sub-argument). They propose exploring variants of generalized rebut, such as requiring the attacker to attack all or some of the sub-conclusions. This is seen as a promising avenue for future research.

Summary updated

BibTeX
@inproceedings{heyninck2017revisiting,
  author = {Jesse Heyninck and Christian Straßer},
  title = {{Revisiting Unrestricted Rebut and Preferences in Structured Argumentation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/151},
}
Figure