Choosing the Lens: Strategic Perspective Activation... | AI Research

Choosing the Lens: Strategic Perspective Activation in Context-Dependent Argumentation
This paper introduces a new way to model how arguments are evaluated in different situations. Often, the same set of arguments is viewed through different "lenses"—such as a financial, technical, or reliability perspective—where the importance of each argument changes depending on the context. The authors propose Context-Dependent Argumentation Frameworks (CDAFs) to formalize this, allowing an agent to strategically choose which perspectives to activate to ensure their preferred argument is accepted.

How the Framework Works

In standard argumentation, the relationship between arguments is fixed. In a CDAF, the authors introduce a "defeat function" that determines which attacks actually succeed based on the current context. This is managed through a "perspective-labeled" system:

• Perspectives: Each argument is assigned a source perspective (e.g., technical or financial).

• Relevance Set ($\rho$): This is the agent’s "action space." The agent chooses which perspectives are active in a given situation.

• Priority ($\pi$): This represents the institutional rules or rankings that the agent cannot change.
  When an agent deactivates a perspective, the arguments associated with it lose their ability to attack others, but they remain present in the system. This asymmetry allows an agent to silence specific "friendly-fire" attacks while keeping the arguments themselves available for support.

Strategic Advantages

The authors demonstrate that this approach is more flexible than existing models, such as Value-Based Argumentation Frameworks (VAFs). In a VAF, an audience can reorder the importance of values, but they cannot turn them off entirely. The paper provides a worked example showing that an agent can achieve their goal by deactivating a perspective—a move that is mathematically impossible to replicate using a VAF audience. By choosing the right "lens," an agent can break structural traps where an argument is otherwise blocked by its own defenders.

Complexity and Decision Problems

The authors define a decision problem called ACTIVATION-MANIPULATION, which asks whether there exists any combination of active perspectives that leads to the acceptance of a target argument.

• For preferred and stable semantics, the problem is NP-complete, meaning it is computationally challenging but solvable by guessing a valid configuration and verifying it.

• For grounded semantics, the problem is at least P-hard. The authors note that these bounds rely on the requirement that the agent must choose a non-empty set of perspectives; if the agent could choose to deactivate everything, the system would become trivial.

Future Directions

While this paper establishes the core logic of perspective activation, several questions remain open. The authors suggest that future research should explore multi-agent scenarios, where different parties might have conflicting goals regarding which perspectives to activate. Additionally, they identify the need for further study into "constrained" environments, such as scenarios where certain perspectives are mandatory or where activating a perspective carries a specific cost. These extensions could provide a more robust framework for real-world applications, such as automated decision-making or agent-based memory systems.