In classical abstract argumentation, reasoning is modeled as a directed graph where nodes represent arguments and edges represent "attacks." If argument A attacks argument B, and A is accepted, then B must be rejected. While this captures basic conflict, it fails to model how humans actually debate. In real-world discourse, we don't just challenge an opponent's premises; we challenge the relevance or validity of their critiques.
To bridge this gap, researchers have developed recursive attack models. These frameworks allow attacks to target not just arguments, but the attack relations themselves.
The Limitation of Atomic Attacks
Traditional models treat attacks as static, indisputable facts. As noted by Amgoud et al., Doder, and Lagasquie-Schiex in Higher-Order Argumentation Frameworks: Principles and Gradual Semantics, this prevents the representation of meta-level critiques. For example, in a legal setting, a lawyer might not dispute a witness's testimony (Argument A) but might argue that the testimony is inadmissible due to a procedural error (Argument C). Here, C does not attack A; it attacks the relationship between A and the verdict.
Without recursive structures, formal systems often reach stalemates in complex debates. Cerutti et al. demonstrate that by allowing "attacks on attacks," logic frameworks can resolve these deadlocks, providing a more decisive evaluation of human discourse by neutralizing specific logical links rather than just the arguments they connect.
Formalizing Meta-Argumentation
The transition to recursive modeling requires a fundamental shift in how we define the "status" of an argument. In the Argumentation Framework with Recursive Attacks (AFRA) introduced by Baroni et al., an attack is treated as a first-class citizen that can be accepted or rejected just like an argument.
This recursive structure allows for several sophisticated reasoning patterns:
- Neutralizing Conflicts: A third factor can override a conflict between two arguments. Cerutti and Cerutti explain that this effectively models how overarching values or goals can resolve specific objections in decision-making.
- Preference-Based Resolution: Often, we resolve conflicts by deciding which argument is "stronger." Modgil proposes Extended Argumentation Frameworks (EAFs), where meta-arguments function as preferences. In this view, arguing that "Argument A is preferred to Argument B" is equivalent to an attack on the attack from B to A.
- Hierarchical Disputes: Reasoning can be seen as a recursive game. Baker et al. and Ginsberg describe this as a dispute between a believer and an unbeliever, where each refutation must move up a priority hierarchy to remain valid.
Integrating Features and Constraints
As these frameworks become more expressive, they must also handle the metadata associated with real-world information. Alfano et al. introduced Featured Argumentation Frameworks (FAF) to incorporate argument-specific characteristics, such as the reliability of a source or the time a claim was made.
Furthermore, the authors of Preferences and Constraints in Abstract Argumentation argue that binary "true/false" logic is insufficient. By allowing three-valued preferences (true, false, or undecided), recursive frameworks can model nuanced agent beliefs, such as preferring that a specific argument remains undecided rather than being proven false.
Computational Feasibility
A primary concern with increasing expressiveness is computational complexity. However, recent work suggests that recursive frameworks can be implemented efficiently. Alfano et al. show that even Recursive Bipolar Argumentation Frameworks (Rec-BAFs), which include both recursive attacks and supports, can be mapped to logic programs without increasing the underlying complexity of standard, non-recursive models.
This formal robustness opens the door to hybrid AI systems. Garcez et al. have explored translating these frameworks into neural networks. By doing so, backpropagation can be used to resolve logical circularities and learn the relative strengths of arguments from data, combining the interpretability of symbolic reasoning with the learning power of connectionist systems.
Key Foundational Work
A Theorem Prover for Prioritized Circumscription (1989) — contributed 1 concept elements to this direction.
Value-based Argumentation Frameworks as Neural-symbolic Learning Systems (2005) — contributed 1 concept elements to this direction.