When humans communicate, they rarely operate like classical theorem provers. In formal logic, if a premise is true, any logically weaker statement derived from it is also indisputably true. But human language and reasoning do not follow these rigid boundaries. As noted in the foundational work Natural, natural language understanding requires moving beyond strict logical structures to accommodate context-dependency, degrees of truth, and varying levels of certainty.
To bridge the gap between human communication and computational systems, researchers study pragmatic inference and contextual entailment—frameworks that treat human "deviations" from formal logic not as cognitive errors, but as structured, context-dependent reasoning.
The Failure of Strict Deduction
Consider how humans evaluate truth in everyday discourse. In formal logic, if a statement is true, a weaker version of that statement is also true. However, Aljoscha Burchardt et al. point out that human textual entailment is governed by pragmatic rules that can actively block these logically valid inferences.
For example, if a text asserts a definitive fact, a human reader will often reject a logically weaker statement of possibility derived from it. This is because, in human conversation, introducing terms of possibility ("It is possible that...") pragmatically implies a degree of uncertainty that contradicts the original text. Because human inference relies heavily on conversational norms and background knowledge, strict deductive logic frequently fails to predict how humans interpret text.
This divergence creates significant vulnerabilities in modern artificial intelligence applications, such as automated fact-checking. If an AI system evaluates a claim based solely on formal logical soundness, it risks misclassifying misleading information.
As argued by Jason Chan et al., logical soundness is an insufficient criterion for neurosymbolic fact-checking with Large Language Models (LLMs). A statement can be entirely valid from a strict deductive standpoint, yet highly misleading to a human reader because it triggers unintended pragmatic inferences. Rather than trying to eliminate these human tendencies, the authors suggest leveraging LLMs' human-like reasoning patterns as a feature to identify when a logically sound statement might still deceive a reader.
Rationalizing the "Illogical"
If human reasoning frequently violates classical logical norms—by either suppressing valid inferences or endorsing invalid ones—how do we model it computationally? Historically, as discussed in Logic Modelling of Cognitive, it has been challenging to explain why individuals make seemingly illogical inferences, often attributing them to faulty cognitive strategies or misinterpretations.
A key insight of this research direction is that these deviations are often internally consistent when evaluated through non-classical frameworks. Christian Eichhorn 0001 et al. demonstrate that many perceived cognitive errors are actually rational behaviors based on alternative plausibility rankings of possible outcomes. By utilizing conditional logic frameworks like Ordinal Conditional Functions (OCFs), the authors show that human inference patterns can be reverse-engineered to reveal the implicit background knowledge and contextual assumptions that make their conclusions rational.
Classical Logic: [Premise] ──────────────────────────► [Strict Deduction]
│ (Can be misleading)
▼
Pragmatic Logic: [Premise] + [Context/Norms] ────────► [Plausible Inference]
Implementing Contextual Entailment
To build AI systems that simulate human-like holistic reasoning, computational frameworks must shift from universal truths to localized, narrative-bound relationships. Early foundations for this shift were proposed by Alain Berrendonner et al., who designed inferential systems capable of handling plausible, context-dependent conclusions rather than relying solely on strict deduction.
In modern natural language processing, this is operationalized through contextual entailment. In machine reading comprehension, for instance, Rajkumar Pujari et al. introduced a method that models the logical relationships between answer choices based entirely on the specific narrative context provided, rather than universal truths. Enforcing these localized pragmatic constraints allows models to capture nuanced dependencies that are only valid within the scope of a given text.
When dealing with the fluid and uncertain nature of common sense, strict logical inference can be replaced by more flexible architectures. Loizos Michael et al. propose an argumentation-based reasoning framework as a more suitable approach for handling the nuances of human interaction. This allows computing systems to apply knowledge based on its contextual relevance and strength, mirroring the flexible, pragmatic way humans navigate the world.
Key Foundational Work
Logic Modelling of Cognitive Reasoning (1983) — contributed 2 concept elements to this direction.
Natural Semantics in Artificial Intelligence (1973) — contributed 1 concept elements to this direction.