Introduction
In the field of formal argumentation—a cornerstone of AI for legal reasoning and multi-agent systems—Dung’s classical framework has long been the standard. However, this model typically treats arguments as individual nodes in a graph where binary edges represent one-on-one attacks. Real-world reasoning is often more complex: a single argument may be insufficient to defeat a claim, requiring a collective set of premises to form a valid contradiction.
This paper, presented at AAAI-20, introduces the nh-categoriser, a novel ranking-based semantics designed for "hypergraph" argumentation frameworks. Unlike traditional models that merely classify arguments as "accepted" or "rejected," the nh-categoriser provides a granular strength score, accounting for joint attacks where sets of arguments work together.
Modeling Joint Attacks with Hypergraphs
The authors define a Hypergraph Argumentation Framework (AS) where an attack is a relationship between a set of arguments and a single target. This allows for a more natural representation of logical dependencies. For instance, if a conclusion requires two premises to be true, an attack on that conclusion might only be successful if both premises are jointly challenged.
The nh-categoriser: A "Weakest Link" Approach
To evaluate the strength of arguments in these complex structures, the authors proposed the nh-categoriser. The core of this semantics is the "weakest link" principle: the effective strength of a joint attack is limited by the weakest argument in the attacking set.
Mathematically, the score of an argument $a$ is calculated recursively based on the scores of the sets attacking it, using a min function to aggregate the strength of each set. Using Brouwer’s fixed-point theorem, the authors rigorously proved that this recursive function always converges to a unique, stable set of scores for any given graph, ensuring the algorithm's reliability for AI implementations.
Formal Evaluation and Postulates
A significant contribution of this work is the generalization of 18 "desirable properties" (postulates) from binary rankings to the hypergraph setting. These include:
- Void Precedence: Arguments that are not attacked should receive the highest possible rank.
- Counter-transitivity: If the attackers of argument $a$ are "stronger" than the attackers of argument $b$, then $b$ should be ranked higher than $a$.
The nh-categoriser was tested against these benchmarks to determine its logical consistency.
As shown in the evaluation table, the nh-categoriser fails Cardinality Precedence—the idea that more attackers always result in a weaker argument. Instead, it prioritizes the quality (strength) of the attackers over raw quantity, a design choice that aligns with the "weakest link" philosophy.
Conclusion and Future Directions
The nh-categoriser represents a "proper generalization" of existing theory; when applied to a simple binary graph, it behaves identically to the well-known h-categoriser. This ensures backward compatibility while expanding the reach of formal argumentation to more sophisticated logical structures.
Future research may explore alternative aggregation functions, such as "synergistic" models where the combined strength of a set exceeds its individual parts. Additionally, incorporating "support" relationships (bipolar frameworks) remains a key frontier for bridging the gap between formal AI models and the nuances of human debate.
Summary updated