Logical Fallacy Detection

Decomposes text into premises and conclusions to evaluate logical form. The system identifies the underlying warrant connecting evidence to a claim.

• Uses discourse parsing to segment text into argumentative units
• Classifies propositions as premise, conclusion, or backing statement
• Detects missing enthymemes where a premise is left implicit
• Example: Identifying that 'We should trust her plan because she's a successful CEO' contains an unstated premise equating business success with planning competence

Maps detected reasoning flaws to a structured ontology of fallacies. Systems typically recognize 20-50 distinct fallacy types across formal and informal categories.

• Formal fallacies: Errors in deductive logical structure (affirming the consequent, denying the antecedent)
• Informal fallacies: Context-dependent errors like ad hominem, straw man, false dilemma, slippery slope
• Appeal fallacies: Appeal to authority, emotion, popularity, or tradition
• Multi-label classification handles arguments containing multiple simultaneous fallacies

Evaluates whether premises are genuinely relevant to the conclusion or merely distract. This requires world knowledge and common sense reasoning beyond surface-level text patterns.

• Distinguishes legitimate expert appeals from fallacious appeal to authority
• Identifies red herrings by measuring semantic distance between premise and conclusion topics
• Uses knowledge graphs to verify whether cited authorities have genuine domain expertise
• Example: Citing a Nobel physicist on climate policy carries different weight than citing them on particle physics

Detects linguistic markers and rhetorical structures commonly associated with fallacious reasoning. Models learn from annotated corpora of debate transcripts, social media arguments, and persuasive essays.

• Lexical cues: 'Everyone knows', 'Obviously', 'Common sense dictates' often signal bandwagon or assertion fallacies
• Syntactic patterns: False equivalence constructions ('X is just like Y'), false dilemma framing ('Either you're with us or against us')
• Pronoun analysis: Overuse of 'they' and 'them' can indicate ad hominem targeting or othering
• Transformer-based classifiers fine-tuned on datasets like LOGIC and Argotario achieve 80%+ F1 scores on fallacy detection benchmarks

Generates human-readable justifications for why a specific passage contains fallacious reasoning. This is critical for educational applications and content moderation transparency.

• Highlights the specific text span containing the fallacy
• Names the fallacy type with a plain-language definition
• Explains why the reasoning fails: 'This argument attacks the person making the claim rather than addressing the claim itself'
• Provides counter-examples showing how a non-fallacious version of the argument would be structured
• Aligns with explainable AI (XAI) principles for content moderation systems

Recognizes that fallacy manifestations vary significantly across domains. A fallacy in legal argumentation looks different from one in political rhetoric or scientific discourse.

• Legal domain: Detects circular reasoning in statutory interpretation, false equivalence in case law citations
• Scientific domain: Identifies correlation-causation confusion, hasty generalization from small sample sizes
• Political domain: Flags whataboutism, tu quoque, and false equivalence in debate transcripts
• Requires domain-specific fine-tuning with annotated examples from each field
• Cross-domain generalization remains an active research challenge in NLP fallacy detection

The computational analysis of discourse to extract the premises, conclusions, and argumentative structures that underpin persuasive text. Before detecting fallacies, systems must first identify the argument components and their relationships.

• Extracts premise-conclusion pairs from raw text
• Identifies argumentative discourse units (ADUs)
• Maps support and attack relationships between claims
• Essential preprocessing step for fallacy classification

A foundational task where a model determines whether a hypothesis can be logically inferred as entailment, contradiction, or neutral from a given premise. NLI provides the reasoning backbone for identifying when conclusions don't follow from their stated premises.

• Classifies logical relationships between text pairs
• Powers formal fallacy detection like non sequitur identification
• Benchmarked on SNLI and MultiNLI datasets
• Distinguishes semantic validity from factual accuracy

The identification of manipulative communication techniques designed to influence opinion through emotional appeal and biased framing rather than objective logic. Propaganda detection overlaps significantly with fallacy detection but focuses on intentional manipulation.

• Identifies techniques like bandwagon, whataboutism, and loaded language
• Uses the 18-technique taxonomy from SemEval-2020 Task 11
• Analyzes both span-level techniques and document-level intent
• Complements fallacy detection in disinformation pipelines

The computational task of determining the attitude or position of a text author towards a specific target claim, typically classified as agree, disagree, or neutral. Stance detection helps contextualize whether a fallacious argument is being deployed in support of or opposition to a claim.

• Operates independently of the author's factual correctness
• Critical for mapping debate structures and echo chambers
• Uses target-specific attention mechanisms
• Distinguishes stance from sentiment and opinion mining

A verification framework that produces human-readable justifications and evidence provenance alongside a veracity label. When integrated with fallacy detection, it explains not just that a claim is false, but which reasoning error produced the invalid conclusion.

• Generates natural language rationales for verdicts
• Links detected fallacies to specific argument components
• Provides auditable reasoning trails for editorial review
• Combines veracity prediction with justification production

The systematic evaluation of fallacy detection models against inputs specifically designed to cause misclassification or bypass detection. Sophisticated disinformation actors craft arguments that exploit model blind spots in logical reasoning.

• Tests against paraphrased fallacies that preserve logical flaws
• Evaluates cross-domain generalization from political to scientific text
• Probes for lexical overfitting to fallacy-indicative keywords
• Measures robustness to implicit vs. explicit fallacy formulations