Human Conditional Reasoning in Answer Set Programming

📝 Paper Summary

Answer Set Programming (ASP) Commonsense Reasoning Cognitive Modeling

This paper introduces new types of program completion in Answer Set Programming to model human pragmatic inferences like Affirming the Consequent and Denying the Antecedent, which are logically invalid but cognitively common.

Core Problem

Standard logic programming and Answer Set Programming (ASP) perform deductive inference but cannot naturally model common human reasoning patterns like Affirming the Consequent (AC) and Denying the Antecedent (DA), which are logical fallacies but pragmatic necessities in daily life.

Why it matters:

Humans often treat 'if' as 'if and only if' (conditional perfection) in daily communication, inferring reverse causality that standard logic misses.
AI systems interacting with humans need to reason in 'cognitively adequate' ways—mirroring human inferences—rather than returning 'unknown' for pragmatically valid conclusions.
Existing approaches like Abductive Logic Programming or Clark's completion only partially cover these patterns and lack a unified framework for all four conditional inference types (AA, AC, DA, DC).

Concrete Example: Given 'If the team wins, they advance' and 'The team advances', standard logic cannot conclude 'The team won' (Affirming the Consequent). Similarly, given 'If the team wins, they advance' and 'The team did not win', logic cannot conclude 'They did not advance' (Denying the Antecedent). Humans frequently make these inferences.

Key Novelty

Conditional Completion in ASP

Introduces weak and strong versions of 'AC completion', 'DA completion', and 'DC completion' to transform standard ASP rules into forms that support pragmatic inferences.
Models 'conditional perfection' (interpreting 'if' as 'iff') by automatically generating converse rules (e.g., inferring antecedent from consequent) within the ASP framework.
Provides a modular mechanism where specific completion strategies can be selectively applied to rules to match different human reasoning contexts.

Evaluation Highlights

Formal characterization of Byrne's suppression task, correctly modeling how humans suppress fallacies when additional context is provided.
Formal correspondence established between the proposed ASP completions and established cognitive science theories like Mental Logic and Conditional Logic.
Demonstrates that Strong AC completion realizes both Affirming the Consequent and Denying the Antecedent simultaneously, mirroring 'conditional perfection' in human language.

Breakthrough Assessment

6/10

A solid theoretical contribution bridging logic programming and cognitive science. It formalizes pragmatic inference in ASP, but relies on theoretical proofs rather than empirical benchmarks or large-scale experiments.

⚙️ Technical Details

Problem Definition

Setting: Modeling conditional reasoning within General Extended Disjunctive Programs (GEDP) under answer set semantics.

Inputs: A logic program Π containing rules r: H ← B and a set of facts/observations.

Outputs: Answer sets of the transformed program Π' representing valid pragmatic inferences (AC, DA, DC).

Pipeline Flow

Input Program (ASP rules)
Completion Transformation (AC, DA, or DC completion applied to rules)
Answer Set Solver (computes stable models of transformed program)

System Modules

Completion Transformation

Augment the original program with additional rules that encode the pragmatic inferences (e.g., adding converse rules for AC).

Model or implementation: Rule-based transformation algorithms (Weak/Strong AC, DA, DC)

ASP Solver

Compute answer sets of the completed program to derive conclusions.

Model or implementation: Standard ASP Solver (conceptually)

Novel Architectural Elements

Formal definition of 8 types of completion (Weak/Strong × AC/DA/DC/Combinations) specifically for GEDP rules.
Introduction of converse rules using default negation to model the non-monotonic nature of human pragmatic inference.

Modeling

Base Model: Answer Set Programming (Logic Programming under Stable Model Semantics)

Comparison to Prior Work

vs. Clark's Completion: Clark's completion produces DA results but cannot handle explicit negation or disjunction in the same way; this paper's DA completion generalizes it.
vs. ALP: ALP models AC as finding hypotheses; this paper models AC as a transformation of the program itself, treating it as a form of deduction in the transformed program.
vs. Stenning and Lambalgen: They focus on specific tasks like suppression; this paper provides a general framework for all four inference types (AA, AC, DA, DC) within standard ASP.
+ 2 more
vs. Pereira et al. (1991) [not cited in paper]: Pereira used reductio ad absurdum for DC; this paper uses explicit transformation rules.
Novel contribution: Provides a unified ASP framework where AC, DA, and DC are all realized via program transformation, allowing modular application to specific rules.

Limitations

Computational complexity of computing answer sets for GEDPs is high (Σ_2^P - complete), potentially limiting scalability.
The approach assumes a specific logical form for conditionals, which may not capture all nuances of natural language pragmatics.
No empirical evaluation with human subjects or performance benchmarks; purely theoretical characterization.
Requires manual or heuristic selection of which completions to apply to which rules (context-sensitivity is discussed but not fully automated).

Reproducibility

The paper is theoretical. It provides formal definitions, theorems, and proofs. No code or dataset is required for reproduction, only verification of the logical derivations.

📊 Experiments & Results

Evaluation Setup

Theoretical analysis and case studies of canonical problems in cognitive science.

Benchmarks:

Byrne's Suppression Task (Cognitive Reasoning Task)
Wason Selection Task (Cognitive Reasoning Task)

Metrics:

Logical correspondence
Satisfiability of inference patterns
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
The paper demonstrates the validity of its approach through formal proofs and case studies rather than numeric benchmarks. The 'results' are the successful modeling of reasoning behaviors.
Byrne's Suppression Task	Inference Pattern Match	Fails to suppress fallacies	Correctly suppresses fallacies	Qualitative improvement

Main Takeaways

Strong AC completion creates a tight coupling (bi-conditional) between antecedent and consequent, enabling both AC and DA inferences.
Weak completions allow for pragmatic inferences without enforcing them, useful for default reasoning where exceptions might exist.
The framework can successfully model non-monotonic behaviors in human reasoning, such as the suppression of fallacies when new information is presented.
The proposed completions naturally extend to handle commonsense reasoning tasks like abduction and counterfactuals within ASP.

📚 Prerequisite Knowledge

Prerequisites

Answer Set Programming (ASP)
Propositional Logic (Modus Ponens, Modus Tollens)
Non-monotonic reasoning
Clark's Completion

Key Terms

AC (Affirming the Consequent): Inferring the antecedent φ from the consequent ψ in a conditional 'if φ then ψ'. Logically invalid but common in human reasoning.

DA (Denying the Antecedent): Inferring ¬ψ from ¬φ in a conditional 'if φ then ψ'. Logically invalid but common in human reasoning.

DC (Denying the Consequent): Modus Tollens: Inferring ¬φ from ¬ψ. Logically valid but not natively supported in standard logic programming deduction.

AA (Affirming the Antecedent): Modus Ponens: Inferring ψ from φ. The standard deduction mechanism.

GEDP: General Extended Disjunctive Program—a logic program allowing disjunction and dual negation (default and explicit).

Conditional Perfection: The pragmatic phenomenon where 'if' is interpreted as 'if and only if' (bi-conditional).

Answer Set: A stable model of a logic program representing a possible set of beliefs or truths.

Weak Completion: A transformation that allows pragmatic inferences (like AC) without forcing them; they become possible but not necessary.

Strong Completion: A transformation that forces the pragmatic inference to hold (similar to a bi-conditional definition).