Choosing the Lens: Strategic Perspective Activation in Context-Dependent Argumentation

Technical Summary: Context-Dependent Argumentation Frameworks and Strategic Activation

Problem Definition
The paper addresses a gap in standard formal argumentation theory regarding the evaluation of arguments under varying external regimes. In real-world scenarios, such as architectural reviews, the same set of arguments and their conflicts remain fixed, but the "lens" through which they are evaluated changes based on the reviewer's priorities (e.g., performance vs. reliability vs. cost). Standard Dung frameworks (1995) lack a mechanism to model this regime selection. While Value-Based Argumentation Frameworks (VAFs) (Bench-Capon 2003) allow for different defeat patterns via audience preferences, they are limited: an audience can only rerank values, not deactivate them entirely. Consequently, VAFs cannot capture scenarios where an agent strategically chooses to exclude specific perspectives from the evaluation process to achieve a desired outcome.

Methodology: Context-Dependent Argumentation Frameworks (CDAFs)
The authors introduce Context-Dependent Argumentation Frameworks (CDAFs) as an extension of Dung's theory. A CDAF is defined as a tuple $\langle A, R, C, \delta \rangle$, where:

• $A$ is a finite set of arguments.
• $R \subseteq A \times A$ is the attack relation.
• $C$ is a finite set of contexts.
• $\delta: C \times R \to \{0, 1\}$ is a defeat function that determines, for each context, which attacks succeed.

The paper focuses on a specialized Perspective-labeled CDAF, where the defeat function is derived rather than given. Arguments are assigned source perspectives ($src: A \to \Pi$). A context is defined by:

1. Relevance Set ($\rho$): A subset of active perspectives ($\rho \subseteq \Pi$). This represents the agent's action space or "lens."
2. Priority Function ($\pi$): A ranking of perspectives within the context.

An attack $(a, b)$ succeeds in a context if and only if the attacker's perspective is active ($src(a) \in \rho$) and the attacker's priority is greater than or equal to the defender's ($\pi(src(a)) \geq \pi(src(b))$). Crucially, this creates an asymmetry: deactivating a perspective silences all attacks mounted by arguments from that perspective, but those arguments remain present in the framework and can still be defended or accepted if they are not attacked by active perspectives.

Key Contributions and Results

1. Strategic Activation: The paper demonstrates that varying the relevance set $\rho$ (activating/deactivating perspectives) can yield acceptance outcomes that are impossible to achieve in VAFs.

   • Worked Example: The authors construct a scenario with arguments $\{a, b, t, d\}$ and perspectives $\{\alpha, \beta, \gamma\}$. The target argument $t$ is structurally trapped: under full relevance (all perspectives active), $t$ is rejected because its defender $b$ also attacks $t$. However, by deactivating perspective $\alpha$ (setting $\rho = \{\beta, \gamma\}$), the "friendly fire" attack $(a, t)$ is silenced. This allows $t$ to be accepted via a preferred extension, a result unattainable under any injective priority in a full-relevance setting.
   • VAF Limitation: The paper proves that the defeat pattern induced by partial activation (e.g., $\rho = \{\gamma\}$, where mutual attacks between $a$ and $b$ both fail) cannot be realized by any VAF audience, as VAF audiences require a strict total order on values, ensuring at least one direction of a mutual attack succeeds.

2. Decision Problem: ACTIVATION-MANIPULATION: The authors formalize the strategic question as a decision problem: Given a framework, a fixed priority, and a target argument $t$, does there exist a non-empty subset $\rho \subseteq \Pi$ such that $t$ is credulously $\sigma$-accepted?

   • Complexity Bounds:
     • For Grounded ($gr$), Stable ($stb$), and Preferred ($pref$) semantics, the problem is shown to be in NP. The authors note that while verifying a preferred extension is typically co-NP, credulous preferred-acceptance coincides with credulous admissibility, allowing the problem to be solved by guessing $\rho$ and an admissible set in a single NP computation.
     • For Stable and Preferred semantics, the problem is NP-hard (and thus NP-complete) via reduction from standard Dung framework acceptance.
     • For Grounded semantics, the problem is in P, though it remains open whether the freedom to choose $\rho$ raises the grounded variant to NP-completeness.
   • Constraint: The requirement that $\rho$ be non-empty is essential; an empty $\rho$ would deactivate all attacks, trivially accepting all arguments.

Significance and Scope
The paper claims significance primarily in defining a new formalism that captures a specific strategic lever: the ability to select which regimes (perspectives) govern an argumentation process, distinct from merely re-ranking them. This models scenarios where an agent controls the "forum" or "track" of a debate rather than the internal logic of the reviewers.

The authors maintain a modest scope, explicitly stating that a full study of the framework is beyond the paper's focus. They identify several open directions:

• Determining tight complexity bounds, particularly for the grounded variant.
• Exploring mechanism design where $\rho$ selection is constrained by costs or mandatory inclusions.
• Investigating multi-agent variants where agents jointly determine $\rho$ through cooperative or adversarial dynamics.
• Examining the phenomenon in applied settings, such as multi-perspective agent memory systems.

The work distinguishes itself from Control Argumentation (which alters the structure) and Manipulation in Incomplete Argumentation (which alters the structure), positioning CDAFs as a model where the structure is fixed, but the defeat function is dynamically modulated by the agent's choice of active perspectives.