Abstract
Human reasoning frequently deviates from classical logic, leading to the suppression of valid inferences or the endorsement of fallacies. This paper proposes that such behavior is not necessarily irrational but can be modeled as internally consistent within a framework of conditional logic. By utilizing Ordinal Conditional Functions (OCF) and c-representations, the authors provide a formal methodology to reverse-engineer the implicit background knowledge guiding human inference patterns.
Introduction
In classical logic, rules such as Modus Ponens (MP) are considered the standard for rational thought. However, psychological research consistently demonstrates that humans often fail to follow these rules, particularly when presented with additional context or counterfactual scenarios. For instance, individuals may suppress a valid inference or endorse "invalid" ones like Affirmation of the Consequent (AC).
This research addresses the lack of a formal model that treats these deviations as rational behaviors based on non-classical logic. Rather than analyzing rules in isolation, the authors focus on inference patterns—the collective application or avoidance of multiple rules—to determine if they are supported by a consistent underlying plausibility model.
Methodology
The authors formalize human reasoning using several key tools from knowledge representation:
- Inference Patterns: A pattern is defined as a 4-tuple representing the acceptance or rejection of four rules: Modus Ponens (MP), Modus Tollens (MT), Affirmation of the Consequent (AC), and Denial of the Antecedent (DA).
- Ordinal Conditional Functions (OCF): These ranking functions assign "implausibility" scores to possible worlds. A conclusion is drawn if its verification is more plausible (lower rank) than its falsification.
- Plausibility Constraints: Each inference rule is translated into a constraint on the plausibility of specific world states.
- C-Representations: This inductive approach generates OCFs from a set of conditionals, allowing the translation of desired inferences into systems of linear inequalities.
- Explanation Generator Algorithm: This core contribution takes an observed human inference pattern and outputs the most concise "conditional knowledge base" (a set of "If-Then" rules) that logically produces that pattern.
Key Insights
The study redefines rationality as internal consistency: "Is there any consistent plausibility ranking that justifies this pattern?" Their analysis reveals that out of 16 possible combinations of the four inference rules, only two are mathematically impossible to satisfy, suggesting that most human "errors" are actually logically consistent under different background assumptions.
By using the Explanation Generator, the authors can hypothesize "hidden" rules. For example, in suppression tasks, participants might treat a conditional in its inverse form, effectively making AC and DA valid moves within their mental model. This approach bridges the gap between rigid classical logic and complex Bayesian probabilities using OCFs as a "semi-quantitative" middle ground.
Results and Discussion
The paper maps behavior from landmark psychological studies to specific formal patterns. For example, the Suppression Task (Byrne, 1989) is mapped to the pattern (¬MP, ¬MT, AC, DA), while counterfactual reasoning (Thompson and Byrne, 2002) follows (MP, MT, ¬AC, DA).
For the Suppression Task, the algorithm suggested that participants might operate under implicit beliefs like "If Lisa is in the library, she (usually) has an essay to write," which explains the endorsement of AC and DA. The results demonstrate that when humans use only classically valid rules (MP and MT), their belief set aligns with "material implication" ($A \Rightarrow B$). Deviations occur when the conditional is treated as a more complex relationship.
Limitations and Future Directions
The current model is limited to four basic inference rules and typically models the "majority" behavior of a group. Future work aims to apply this framework to individual reasoners to assess personal internal consistency. Additionally, while the algorithm generates hypotheses for background knowledge, these require further experimental validation to confirm if participants actually hold the predicted "hidden" beliefs.
Summary updated