Justifying Argument Acceptance with Collective Attacks: Discussions and Disputes

Giovanni Buraglio, Wolfgang Dvorák, Matthias König, Markus UlbrichtIJCAI 2024

Abstract

This paper introduces an efficient methodology for justifying argument acceptance using SETAFs—argumentation frameworks that incorporate "collective attacks," where a group of arguments jointly defeats another. By generalizing traditional discussion games and dispute trees to this expressive format, the authors demonstrate that justifications become significantly more concise and human-readable compared to standard binary-attack models, addressing the issue of "argument bloat" in Explainable AI (XAI).

Introduction: The Problem of Argument Bloat

In Formal Argumentation, Abstract Argumentation Frameworks (AFs) are the standard for modeling conflict. However, in binary AFs, attacks are strictly one-to-one (Argument A attacks Argument B). When real-world knowledge bases are translated into these graphs, they often suffer from "argument bloat," where a single logical conflict generates dozens of redundant arguments and attacks.

When an AI explains a conclusion via discussion games (a dialogue between a Proponent and an Opponent) or dispute trees, these explanations in bloated AFs become long, repetitive, and difficult for humans to follow. This lack of conciseness is a major hurdle for XAI in sensitive domains like legal or medical reasoning.

Methodology: SETAFs and Discussion Games

The authors propose shifting the foundation from standard AFs to SETAFs (Argumentation Frameworks with Collective Attacks). In a SETAF, a set of arguments can collectively attack a target, even if no individual member of that set is sufficient to do so alone.

Figure 1: A sample SETAF (SF) illustrating collective attacks, such as the set {d, e} attacking argument b, and {b, c} attacking argument a.

Key methodological contributions include:

  • SETAF Discussion Games: A new protocol where a Proponent (P) defends a target against an Opponent (O). If O claims a set of arguments attacks P's position, P only needs to defeat one member of that set to neutralize the threat.
  • Concise Discussions: To eliminate redundancy, if the Proponent has already defeated an argument $x$, the Opponent is prohibited from using any future collective attacks that rely on $x$.
  • SETAF Dispute Trees: A tree-based visualization of these games, providing a structural map of the justification.

Illustrative Example: The Vacation Scenario

To demonstrate the efficiency of SETAFs, the authors use a vacation-planning scenario (Example 1.1). In this scenario, a traveler (Caroline) chooses between a small village (sv) or an all-inclusive hotel (all).

Figure 2: The SETAF representation of the vacation planning scenario, showing mutual attacks between 'sv' and 'all', and a collective attack on 'af' (affordable).

In a traditional AF instantiation of this problem, a discussion might require 11 moves with significant repetition. In the SETAF version, the same justification is achieved in just 3 moves, directly identifying the "all-inclusive" option as the sole culprit for the conflict regarding affordability.

Key Insights and Results

  • Expressivity Equals Efficiency: By allowing collective attacks, the underlying model captures the "root cause" of a conflict more effectively, leading to simpler explanations.
  • Computational Bounds: The length of a winning strategy for an admissible set $E$ is at most $2 \cdot |H| + 1$, where $H$ is a "hitting set" of the attacks against $E$. This ties explanation length to the number of unique ways to defeat opposition rather than the total number of arguments.
  • Complexity: Finding the shortest winning strategy remains NP-complete, consistent with traditional AFs. Thus, SETAFs provide better explanations without increasing theoretical complexity for optimization.

Conclusion and Future Work

The move to SETAFs offers a promising path for more intuitive XAI. Future research will focus on extending these games to other semantics (like grounded or stable semantics) and conducting human evaluation studies to confirm that collective-attack dialogues are indeed more intuitive for non-experts than binary ones.

Summary updated

BibTeX
@inproceedings{buraglio2024justifying,
  author = {Giovanni Buraglio and Wolfgang Dvorák and Matthias König and Markus Ulbricht},
  title = {{Justifying Argument Acceptance with Collective Attacks: Discussions and Disputes}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure