TL;DR
This paper introduces a novel way to quantify the "intensity" or "harmfulness" of individual attacks within an argumentation system. By adapting the Shapley value, a concept from cooperative game theory, the authors provide a unique and principled method to measure how much each attack contributes to weakening an argument, offering a more nuanced understanding of argumentation dynamics.
What Problem Does This Paper Address?
In artificial intelligence, argumentation frameworks are used to model reasoning and decision-making. These frameworks represent arguments as nodes in a graph, with attacks between arguments depicted as edges. While existing "semantics" can tell us which arguments are generally accepted or rejected, they often don't explain why an argument is weakened. Specifically, they don't reveal how much each individual attack contributes to an argument's reduced acceptability. This lack of insight makes it difficult to understand the true impact of specific attacks, identify redundant ones, or strategize effectively in persuasive dialogues. The paper aims to fill this gap by providing a quantitative measure for the intensity of each attack.
Methods
The core of the paper is the introduction of a "contribution measure" for attacks. This measure takes an argumentation framework (graph + semantics) and assigns a weight to each attack, representing its contribution to the target argument's loss of strength. The authors first define a set of desirable axioms for such a measure, including:
- Anonymity: The identity of arguments shouldn't affect attack contributions.
- Dummy: Attacks that have no impact on an argument's loss should have zero contribution.
- Symmetry: Symmetric attacks (those with similar marginal contributions) should receive equal contributions.
- Dominance: Attacks with greater marginal contributions should receive greater contributions.
- Coherence: An attack's contribution should be consistent across different semantics if its marginal contribution is the same.
They then prove a key theorem: the Shapley value is the unique measure that satisfies these axioms. The Shapley value, originally developed for cooperative games, is adapted here. For an attack r targeting argument a, its contribution is calculated as a weighted average of its marginal contribution to the loss of a across all possible subsets of a's other attackers. This means the measure considers how removing r affects a's loss, taking into account the presence or absence of other attackers.
Datasets
Not discussed in the paper. The paper focuses on theoretical framework development and uses illustrative examples rather than empirical evaluation on datasets.
Key Insights and Ideas
- Quantifying Attack Intensity: The paper moves beyond simply knowing if an argument is attacked to understanding how much each attack matters. This is crucial for deeper analysis and strategic reasoning.
- Shapley Value as a Universal Measure: The surprising and powerful result is that the well-established Shapley value, when applied to the "game" of argumentation attacks, uniquely satisfies a set of intuitive axioms for measuring attack intensity. This provides a theoretically sound foundation for the proposed measure.
- Marginal Contributions are Key: The Shapley value's reliance on marginal contributions highlights that the impact of an attack is not absolute but depends on the context of other attacks. This is a more realistic way to assess harm than simply looking at the source argument's strength.
- Distinguishing "Dummy" Attacks: The axioms help formalize what constitutes a "dummy" attack (one with no real impact), which is essential for pruning irrelevant information in complex argumentation systems.
Key Results
- The paper formally defines "contribution measure" for attacks and proposes five axioms that such a measure should satisfy.
- It proves that the Shapley value is the unique contribution measure satisfying these axioms.
- The paper demonstrates how the Shapley value can be applied to various existing argumentation semantics (e.g., Dung's extension semantics, h-categorizer semantics) and analyzes its properties under these semantics.
- The results show that the Shapley measure can distinguish between attacks that are truly harmful and those that are not, and that its behavior is consistent with the underlying semantics.
Limitations and Future Work
- Computational Complexity: While the Shapley value provides a theoretically sound measure, its direct computation can be computationally expensive, especially for arguments with many attackers. The paper doesn't delve deeply into efficient computation methods.
- Semantics Dependence: While the Shapley value is proposed as a universal measure, its output (the specific contribution values) is dependent on the chosen argumentation semantics. The paper explores this dependence but acknowledges that different semantics can lead to different attack intensities.
- Future Work: The authors suggest extending this work to other argumentation semantics, including probabilistic and weighted argumentation frameworks, and exploring more efficient computational approaches. They also mention investigating the properties of the Shapley measure in settings with different types of argumentation graphs.
Summary updated