Satisfying Rationality Postulates of Structured Arg... | AI Research

Satisfying Rationality Postulates of Structured Argumentation Through Deductive Support – Technical Report
This paper introduces Deductive ASPIC$^{\ominus}$, a new framework designed to improve how artificial intelligence systems reason through structured arguments. In AI, structured argumentation combines logical rules with abstract frameworks to determine which arguments are acceptable. A major challenge in this field is ensuring that these frameworks remain logically consistent and robust. The authors propose a solution that successfully meets five critical "rationality postulates"—standards that ensure an argumentation system behaves logically and predictably—even when using flexible, credulous reasoning methods.

Bridging Existing Frameworks

Previous attempts to satisfy these five rationality postulates—closure, direct consistency, indirect consistency, non-interference, and crash-resistance—have often struggled to do so simultaneously. Specifically, earlier models either ignored certain types of attacks (undercuts) or failed to maintain logical consistency when using preferred semantics. Deductive ASPIC$^{\ominus}$ addresses these gaps by merging two existing approaches: "gen-rebuttals," which allow for more complex attacks between arguments, and "Joint Support Bipolar Argumentation Frameworks" (JSBAFs), which track how strict logical rules support conclusions. By combining these, the framework creates a more unified and sound logical structure.

How the Approach Works

The core of the framework is the use of "deductive support." In this system, arguments are built from strict rules (which are exceptionless) and defeasible rules (which allow for exceptions). The framework tracks not only how arguments attack one another but also how they support each other through strict logical deduction.
To ensure the system remains logical, the authors define "legal labelings." These labels (IN, OUT, or UNDEC) determine whether an argument is accepted, rejected, or undecided. The rules for these labels are designed to propagate the rejection of arguments in a way that mirrors classical logical principles, such as contraposition. This ensures that if the premises of a strict rule are accepted, the conclusion is also accepted, satisfying the closure postulate.

Key Results and Robustness

The authors demonstrate that Deductive ASPIC$^{\ominus}$ is the first framework of its kind to satisfy all five rationality postulates under preferred semantics while accounting for both rebuttals and undercuts. By incorporating preferences—which help resolve conflicts between competing arguments—the system remains flexible without sacrificing logical integrity. The framework also avoids the pitfalls of "restricted rebuttals," which have been criticized for producing counter-intuitive results in dialectical reasoning.

Considerations for Future Research

While this framework provides a significant step forward in building robust argumentation systems, it is presented as a technical foundation. The authors note that this work opens new avenues for exploring how these sound logical principles can be applied to more complex, real-world AI reasoning tasks. By successfully balancing the need for flexible, credulous reasoning with the strict requirements of formal logic, this research provides a reliable basis for future developments in non-monotonic reasoning and knowledge representation.