TL;DR
This paper proposes a unified theory of argumentation that defines argument acceptability precisely. It demonstrates that logic programming and nonmonotonic reasoning are essentially different forms of argumentation, offering a general method for building argumentation systems. This work bridges the gap between symbolic reasoning and logic programming, providing a foundational framework for more robust AI reasoning systems.
What Problem Does This Paper Address?
Humans excel at argumentation – constructing, defending, and attacking claims. This ability is fundamental to intelligence, enabling us to navigate complex situations, express opinions, and reason scientifically. While AI has made strides in various reasoning paradigms like nonmonotonic reasoning and logic programming, a unified understanding of how these approaches relate to human-like argumentation has been lacking. Specifically, the paper addresses the challenge of defining argument acceptability – when should we believe an argument? – and how this concept underpins different AI reasoning formalisms. The authors aim to bridge the gap between the intuitive, human-centric notion of argumentation and the formal, often disparate, methods used in AI for knowledge representation and reasoning.
Methods
The core of the paper is the introduction of a formal theory of argumentation frameworks. An argumentation framework is defined as a pair <AR, attacks>, where AR is a set of arguments and attacks is a relation specifying which arguments attack others.
Key concepts introduced include:
- Conflict-free sets: A set of arguments where no argument attacks another within the set.
- Admissible sets: A conflict-free set
Swhere every argument attacking an argument inSis itself attacked by an argument inS. This captures the idea of defending an argument against counterarguments. - Preferred extensions: Maximal admissible sets. These represent the "credulous" view of what a rational agent might accept.
- Grounded extension: The least fixed point of a characteristic function that iteratively determines acceptable arguments. This represents the "skeptical" view, what can be definitively proven.
- Stable extensions: Conflict-free sets that attack every argument not in the set.
The paper then demonstrates how various AI reasoning approaches, including Reiter's default logic, Pollock's defeasible logic, and logic programming (specifically stable and well-founded semantics), can be reformulated as instances of these argumentation frameworks. A novel contribution is a general method for generating metainterpreters for argumentation systems, viewing them as a special form of logic programming with negation as failure.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Argumentation as a Unifying Principle: The paper's most significant insight is that many seemingly different AI reasoning formalisms (nonmonotonic reasoning, logic programming) are, at their heart, different instantiations of argumentation. This provides a powerful unifying perspective.
- Formalizing "Acceptability": By defining admissible and preferred extensions, the paper offers a rigorous mathematical definition for when an argument should be considered "acceptable" or believable, moving beyond intuitive notions.
- Bridging Logic Programming and Argumentation: The demonstration that logic programming can be viewed as argumentation with negation as failure is crucial. This allows for the development of general metainterpreters, akin to compiler-compilers, for argumentation systems.
- Skeptical vs. Credulous Reasoning: The distinction between grounded (skeptical) and preferred (credulous) extensions elegantly captures different levels of belief or acceptance, mirroring human reasoning where some might be more cautious than others.
Key Results
- Formal Theory of Acceptability: The paper presents a precise theory for defining the acceptability of arguments within argumentation frameworks.
- Argumentation Frameworks for Existing Systems: It shows that logic programming and major nonmonotonic reasoning formalisms are indeed argumentation systems, differing primarily in their argument structures.
- General Metainterpreter Generation: A method is provided for generating metainterpreters for argumentation systems, treating them as a form of logic programming with negation as failure.
- Coincidence of Semantics: For certain classes of frameworks (e.g., well-founded, uncontroversial), the paper proves that different semantics (stable, preferred, grounded) coincide, simplifying reasoning and analysis. For instance, for uncontroversial frameworks, preferred extensions are stable.
Limitations and Future Work
- Single Conflict Type: The current theory primarily handles one type of conflict between arguments (direct attacks). The authors acknowledge that real-world argumentation often involves multiple conflict types, such as reductio ad absurdum and conflicts based on specificity or generality.
- Generalizing to Multiple Conflict Types: Future work is needed to generalize the theory to handle these more complex conflict scenarios. The paper mentions ongoing research in this direction, including work by Bondarenko, Toni, and Kowalski.
- Efficiency: While the paper provides a theoretical foundation, practical implementation and efficiency improvements for argumentation systems are noted as areas for further development, suggesting the application of techniques like partial evaluation and program transformation.
- Knowledge Representation: The authors envision their work as a foundation for developing knowledge representation formalisms that can communicate across different systems, highlighting the importance of handling diverse knowledge representation languages.
Summary updated