Computational Argumentation: Reasoning, Dynamics, and Supporting Explainability

Johannes Peter WallnerIJCAI 2024

Abstract

Computational argumentation provides a formal framework for reasoning with conflicting or incomplete information, mimicking human debate to reach robust conclusions. However, its practical application is hindered by extreme computational complexity and the "exponential explosion" of arguments in structured settings. This paper surveys research into overcoming these barriers through optimized logic solvers (SAT/ASP), structural graph analysis, and faithful abstraction techniques. By bridging the gap between high-level formal logic and practical implementation, this work paves the way for efficient and explainable AI-driven decision support systems.

Introduction

Computational argumentation is a critical subfield of Artificial Intelligence designed to model the process of reaching conclusions from conflicting data. While highly applicable in domains requiring transparent reasoning, such as law and medicine, it faces two primary challenges. First, many reasoning tasks are computationally "hard," often exceeding NP-completeness and sitting high in the polynomial hierarchy. Second, structured argumentation systems can generate an unmanageable number of arguments from a simple set of rules, leading to information overload for both human users and computer systems.

The standard pipeline for computational argumentation, from knowledge base instantiation to drawing conclusions.

Johannes P. Wallner's research agenda addresses these hurdles by identifying structural shortcuts, developing efficient algorithms, and creating methods to simplify complex argument graphs for better explainability. The goal is to make complex reasoning both computationally feasible and human-understandable.

Methodological Framework

The research outlines several sophisticated approaches to tame the complexity of argumentation:

  • SAT and ASP Solving: Rather than building bespoke engines, argumentation problems are translated into Satisfiability (SAT) and Answer Set Programming (ASP) formats. This allows the field to leverage decades of performance optimization in general-purpose logic solvers.
  • Compact Representations: To avoid the "exponential explosion" of arguments, methods were developed to reason over a "core" knowledge base without explicitly constructing every possible argument. This mapping between structured and abstract frameworks is essential for scalability.
  • Structural Restrictions: By identifying specific graph properties—such as acyclicity or bipolarity (where relations are strictly attacks or supports)—the authors demonstrate that complexity can be reduced to manageable, polynomial-time levels.
  • Existential Abstraction: This involves "clustering" related arguments into a single node. The challenge lies in ensuring the abstraction is "faithful," meaning the conclusion reached in the simplified graph remains identical to the one in the original, complex graph.
Mapping structured arguments (A1-A7) to an abstract framework to manage information overload.

Key Insights and Algorithmic Results

A significant finding of this research is that for many formalisms, such as Assumption-Based Argumentation (ABA) and ASPIC+, full argument instantiation is often unnecessary. Reasoning can be performed directly on the underlying rules, matching the lower complexity of abstract frameworks.

Furthermore, while Abstract Dialectical Frameworks (ADFs) offer superior expressive power compared to standard Dung-style frameworks, they come with a "one level higher" cost in computational complexity, requiring nested solver calls. The research has successfully mapped the complexity of various ADF fragments and produced state-of-the-art solvers for structured argumentation. Additionally, the paper introduces the concept of "Dynamic Stability," providing methods to check if adding new information could flip a current result—a vital feature for real-time decision support.

Conclusion and Future Directions

While significant progress has been made in algorithmic efficiency and complexity mapping, the author acknowledges that the field must continue to evolve toward human-centric design. Future work points toward interactive systems where users can "zoom in or out" of an argument graph to understand conclusions at different levels of detail. Developing simplification procedures that help human users digest the core components of a complex debate, alongside better handling of user preferences, remains a promising frontier for computational argumentation.

Summary updated

BibTeX
@inproceedings{wallner2024computational,
  author = {Johannes Peter Wallner},
  title = {{Computational Argumentation: Reasoning, Dynamics, and Supporting Explainability}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure