Abstract
This paper presents a formal investigation into "lifting operators" within structured argumentation—mathematical tools used to rank arguments based on the perceived strength of their individual components. The authors define a comprehensive set of 14 postulates, evaluate existing operators against these properties, and establish a hierarchy of relations between them. Key findings include characterizations of the Elitist and Democratic operators and the identification of logical incompatibilities between certain desirable properties, providing a roadmap for selecting appropriate ranking logics in AI systems.
Introduction
In the field of Artificial Intelligence, structured argumentation allows machines to reason through conflicting information. Arguments are constructed from "defeasible" elements—premises or rules that may be incorrect or superseded. A central challenge is "lifting": taking a preference ranking of individual elements (e.g., "Rule A is more reliable than Rule B") and transforming it into a ranking of sets of elements. While various operators exist, a systematic understanding of their properties and suitability for frameworks like ASPIC+ has been lacking.
Methods
The authors employ an axiomatic approach, surveying lifting operators from argumentation theory, social choice, and utility theory (e.g., Elitist, Democratic, Kelly’s, and Fishburn’s operators). They formulate 14 mathematical postulates that a "good" lifting operator should ideally satisfy, such as Extension (if element $x$ is better than $y$, then the set $\{x\}$ should be better than $\{y\}$). Through formal mathematical proofs, they determine which operators satisfy which postulates and identify "impossibility results" where certain properties are logically incompatible.
Key Insights and Operator Relations
The analysis reveals that Kelly’s operator serves as a foundational baseline; nearly all other rational operators extend it. A major theoretical insight is the proof that it is impossible to simultaneously satisfy Extension, Strong Left Extension, and Strong Right Extension. This forces a design choice between "Elitist" and "Democratic" behaviors.
The authors mapped the relationships between these operators into a clear hierarchy, clarifying how different methods contain or overlap with one another.
Results and Characterizations
The primary result is a comprehensive "satisfaction matrix" (Table 1) that evaluates 10 major operators against the 14 postulates. This serves as a "cheat sheet" for researchers to choose the right logic for their specific application.
The paper also reveals that some popular operators from social choice (like Smyth’s or Fishburn’s) fail the "Reasonable Inducing" property, making them unsuitable for the ASPIC+ framework. Finally, the authors successfully "boxed in" the Elitist and Democratic operators by identifying the exact sets of postulates that uniquely define them.
Conclusion
This study provides the first comprehensive formal map of lifting operators in structured argumentation. By identifying which operators are "rational" under different conditions, the authors provide essential guidance for AI researchers. Future work may explore dynamic preference updates and the alignment of these formal models with human cognitive processes in weighing competing arguments.
Summary updated