Presented at AAAI-21, this research addresses a fundamental limitation in artificial intelligence: the inability of standard logic to handle "if-then" reasoning when exceptions and counterfactual scenarios overlap. The authors propose Contextual Conditional Reasoning, a framework that explicitly incorporates context into conditional statements, allowing AI to distinguish between typical exceptions (e.g., penguins) and purely imaginary scenarios (e.g., flying dodos).
The Problem: Rigid Logic and Counterfactual Blind Spots
Classical logic is often too rigid for real-world reasoning. While non-monotonic reasoning (like the KLM framework) introduced "defeasible" rules—allowing for exceptions like "birds usually fly, but penguins don't"—it struggles with counterfactuals.
If an AI knows dodos are extinct, standard systems might either fail to reason about them at all or fail to distinguish between a "typical exception" (a penguin, which exists but is atypical) and a "counterfactual entity" (a dodo, which does not exist). The goal is a system that maintains coherent expectations about the real world while remaining capable of robust "what-if" reasoning.
Methodology: Tiered Reality and Contextual Parameters
The core innovation is extending standard conditional language ($\alpha \mid\sim \beta$) into a contextual form: $\alpha \mid\sim_\gamma \beta$. This is read as: "In the context of $\gamma$, if $\alpha$ holds, then $\beta$ holds."
Epistemic Interpretations
To support this, the authors developed a semantic framework that ranks possible worlds into two distinct tiers:
- Finite Ranks: Represent plausible worlds or typical expectations (the "real-world" model).
- Infinite Ranks: Represent implausible but logically possible worlds (the "counterfactual" model).
Minimal Closure and Computational Efficiency
The authors define Minimal Closure (MC) as the gold standard for logical entailment in this framework. Crucially, they prove that reasoning within this complex contextual system can be reduced to standard propositional logic checks. This allows the framework to leverage existing, highly optimized SAT solvers, making it computationally viable for practical AI applications.
Key Insights
- Context as a First-Class Citizen: By making context ($\gamma$) an explicit parameter, the logic satisfies key properties from Belief Revision theory (the AGM framework), such as Inclusion and Vacuity.
- Coherent Counterfactuals: The system can simultaneously maintain that "dodos don't exist" in a general context while correctly inferring that "if dodos existed, they would be birds" in a counterfactual context, without causing logical collapse.
- The Mapping of Contexts: The framework uses a mapping system to determine which rank of reality a specific context refers to, ensuring that reasoning remains grounded in the appropriate tier of plausibility.
Results and Future Work
The authors provided a Representation Theorem, proving that their "Full Contextual Conditionals" (FCCs) are exactly characterized by their tiered epistemic interpretations. They also demonstrated that for any consistent set of rules, there exists a unique "best" model (the Minimal Closure).
Future research aims to extend this framework from simple propositional logic to Description Logics, which are essential for the Semantic Web and complex ontologies. Additionally, the authors plan to explore "Lexicographic Closure" variants to enable more nuanced, human-like reasoning that goes beyond the cautious nature of Minimal Closure.
Summary updated