Defeasible Normative Reasoning: A Proof-Theoretic Integration of Logical Argumentation

Ofer Arieli, Kees van Berkel, Christian StraßerAAAI 2024 · DOI

Abstract

This paper introduces a unified framework for handling conflicting norms by integrating proof theory, formal argumentation, and Input/Output (I/O) logics. The authors propose the Annotated Deontic Argumentation Calculus (ADAC), which incorporates status annotations (Accepted, Rejected, Introduced) directly into logical proofs. This approach provides a transparent, explainable system for resolving contradictions in normative reasoning. The study establishes a fundamental threefold equivalence between ADAC, grounded extensions in argumentation, and free consequences in I/O logic, bridging the gap between symbolic reasoning paradigms.

Introduction: Resolving Normative Conflicts

In real-world scenarios, rules and obligations frequently clash. Legal systems may contain overlapping regulations, or autonomous agents may receive contradictory "ought-to-do" commands. Standard logic fails in the presence of such contradictions due to the principle of explosion. While Nonmonotonic Reasoning (NMR) allows new information to invalidate previous conclusions, existing methods are often fragmented. Formal argumentation excels at visualizing conflicts but remains abstract; I/O logics are powerful for modeling norms but lack computational transparency; and proof theory is computationally efficient but typically assumes monotonicity.

The authors address this fragmentation by creating a unified, transparent bridge that explains why a conclusion is accepted or rejected in the face of conflict.

Methodology: The Annotated Deontic Argumentation Calculus (ADAC)

The core innovation is the Annotated Deontic Argumentation Calculus (ADAC), an evolution of standard sequent-style proof systems. ADAC introduces several key technical features:

  1. Status Annotations: Every step in a proof is tagged with a label: [i] for introduced, [a] for accepted, or [r] for rejected.
  2. Internalized Attack Rules: Specific rules (Acpt-1, Acpt-2, Rjct) allow the system to reason about its own derivations. If a proof shows a norm is inapplicable, it can "attack" and reject other derivations relying on that norm.
  3. Labelled Languages: The system uses labels ($f$ for facts, $o$ for obligations, $c$ for constraints) to distinguish the roles of different propositions without complex modal operators.
Argumentation frameworks representing conflicting norms and their resolutions.

The system's behavior is illustrated through a running example involving conflicting academic norms (e.g., funding vs. supervisor promises), demonstrating how "floating conclusions" are handled through skeptical reasoning.

Key Results: A Threefold Correspondence

The paper's primary contribution is the mathematical proof of a "threefold correspondence." This result demonstrates that three distinct reasoning paradigms are mathematically equivalent under specific conditions.

The threefold correspondence between ADAC, I/O logic, and grounded argumentation.

The authors establish three major equivalence theorems:

  • Theorem 1: A formula is "Accepted" in the ADAC proof system if and only if it belongs to the grounded extension of the induced argumentation framework.
  • Theorem 2: Acceptance in ADAC is equivalent to "free consequences" (the intersection of all maximally consistent subsets of norms) in I/O logic.
  • Theorem 3: This completes the triangle, directly linking grounded argumentation extensions to the free consequences of I/O logic.

These results provide a rigorous foundation for Explainable AI (XAI) in domains like law, ethics, and policy, where the rationale behind a decision is as important as the decision itself.

Discussion and Future Work

While the ADAC framework provides a robust foundation for skeptical normative reasoning, it has certain limitations. The current system is strictly "skeptical," meaning it only accepts non-controversial truths. While ideal for safety-critical applications, it does not yet support "credulous" reasoning for exploring multiple valid but conflicting worldviews.

Furthermore, the computational complexity of proof-search in this framework remains to be fully analyzed for large-scale norm bases. Future research could also extend the framework to handle dynamic knowledge bases where norms change or are added during the reasoning process.

Summary updated

BibTeX
@inproceedings{arieli2024defeasible,
  author = {Ofer Arieli and Kees van Berkel and Christian Straßer},
  title = {{Defeasible Normative Reasoning: A Proof-Theoretic Integration of Logical Argumentation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i9.28913},
}
Figure