Logical Fallacy Detection

Detection systems rely on a structured ontology of fallacies adapted for legal discourse. Common categories include:

• Formal Fallacies: Deductive errors like affirming the consequent or denying the antecedent.
• Informal Fallacies: Content-dependent errors such as appeal to authority (citing non-binding dicta as binding) or straw man (mischaracterizing opposing counsel's argument).
• Jurisprudential Fallacies: Domain-specific errors like appeal to novelty (arguing a modern statute overrides a settled common law principle without explicit repeal) or false equivalence between materially distinct precedents.

A robust taxonomy is the prerequisite for any supervised classification approach.

The foundational layer of detection uses lexico-syntactic pattern matching to flag known fallacy indicators:

• Discourse Markers: Phrases like 'it is obvious that', 'everyone agrees', or 'logic dictates' often precede circular reasoning or appeals to common belief.
• Syntactic Templates: Regex and dependency parse patterns capture structures like 'If A then B. B is true, therefore A is true' (affirming the consequent).
• Keyword Triggers: Terms like 'slippery slope', 'red herring', or 'ad hominem' are used to bootstrap retrieval for deeper analysis.

These systems offer high precision but low recall, serving as a first-pass filter before neural evaluation.

Modern systems fine-tune transformer-based models on annotated legal corpora to classify spans of text by fallacy type. The task is framed as:

• Sequence Classification: An entire argument block is labeled with a fallacy category.
• Token-Level Tagging: The exact span containing the fallacious reasoning is identified and tagged.
• Natural Language Inference (NLI): The system tests whether a conclusion can be logically entailed from its stated premises. A lack of entailment, combined with rhetorical confidence, signals a potential non sequitur.

Models are typically initialized from legal pre-trained checkpoints like Legal-BERT or CaseLaw-BERT to capture domain-specific semantics.

Fallacies are often detectable not by the text of a single statement, but by the topology of the argument graph. Detection algorithms analyze:

• Circularity: A cycle in the directed graph where Claim A supports Claim B, which in turn supports Claim A.
• Missing Premises: A conclusion node with no incoming support edges from evidence or legal rules, indicating an enthymeme that may mask a fallacy.
• Isolated Attacks: A rebuttal node that attacks a claim without engaging with its supporting premises, characteristic of a straw man argument.

Graph-based methods complement textual classifiers by catching structural errors invisible to a linear reading.

A critical requirement for legal AI is algorithmic explainability. Fallacy detection systems must not only flag an error but also justify the flag:

• Feature Attribution: Techniques like SHAP or integrated gradients highlight the tokens most responsible for a fallacy classification.
• Rule Grounding: The system maps a detected fallacy to its formal definition in a curated fallacy knowledge base.
• Citation Verification: When a fallacy involves a mischaracterized authority (e.g., straw man or false appeal), the system cross-references the cited source against a ground-truth database to prove the distortion.

This ensures the output is an auditable finding, not an opaque model prediction.

To harden detection models, engineers employ adversarial data augmentation:

• Fallacy Injection: Logically sound arguments are programmatically modified to introduce specific fallacies (e.g., replacing a valid warrant with an appeal to emotion).
• Paraphrasing Attacks: Fallacious arguments are rewritten using diverse syntactic structures and legal jargon to prevent models from overfitting to surface patterns.
• Contrastive Training: Models are trained on pairs of sound and fallacious versions of the same core argument to learn the precise logical boundary.

This approach improves robustness against sophisticated or subtly worded fallacies in real legal briefs.