Translating natural language to first-order logic for logical fallacy detection

Implementation & Models

  • Repository exposes a script taking --model_name (LLM for translation) and --nli_model_name (HuggingFace NLI classifier), but doesn’t ship a pretrained NLI model, causing confusion about how to run it.
  • Some commenters wish this were a fully trained, open model with RL-based translation instead of a wrapper around external models.

Anti‑Propaganda Ambitions vs Reality

  • Several participants are excited about using FOL to expose fallacies, decompose arguments, and help people see inconsistencies in propaganda and value systems.
  • Others argue propaganda relies more on selective framing, omission, emotional salience, and identity than on formal fallacies; logical checking alone cannot capture those.
  • Idea emerges of user‑side “filters” (LLM/FOL layers) that rewrite, annotate, and debias incoming media streams, but this is seen as only one part of a larger solution.

Human Rationality, Values, and Virtues

  • Debate over how rational people are: some say humans largely reason within their value systems; others claim people flex values to preserve identities or idols.
  • Distinction drawn between “values” (evaluated outcomes) and “virtues” (presumed goods). One view: conservative politics especially centers on virtues (e.g., “capitalism vs socialism”), which makes virtue‑framed propaganda powerful and relatively logic‑proof.
  • Others counter that motivated reasoning and denial occur across political tribes and that empirical psychology on asymmetries is fragile.

First‑Order Logic: Capabilities and Limits

  • Some praise FOL as the standard formalism worth learning; others note it’s poor at modeling time, belief, negation, and real human reasoning.
  • Gödel and undecidability are discussed: they don’t forbid proof checking, but they limit universal decision procedures; heuristics still useful.
  • Multiple comments stress that logic can verify consistency relative to a spec but not whether the spec (or moral axioms) is correct—echoing the “formal specification problem.”

Datasets, Semantics, and Practicality

  • Strong criticism of the LOGIC and LOGICCLIMATE benchmarks: examples mislabel tautologies and legitimate causal claims as fallacies and even quote a climate op‑ed selectively to manufacture a “false causality” case.
  • Linguists and NLP veterans argue natural language semantics (Montague grammar, DRT, pragmatics, implicature) are far richer than the paper’s ad‑hoc FOL mappings, so robustness on real text is doubtful.
  • Nonetheless, commenters see niche utility for constrained domains: contracts, laws, technical specs, classroom logic assistants, and pinpointing isolated fallacious sentences in articles.

Relation to LLMs and Informal Fallacies

  • People share prompts already used with current LLMs to extract premises, conclusions, and fallacies and to “steelm​an” arguments; some examples show that state‑of‑the‑art LLMs handle textbook fallacies well.
  • Several note that many real arguments are persuasive, informal, and probabilistic; detecting formal fallacies doesn’t settle truth or persuade opponents and can be waved away as “biased logic.”
  • Informal fallacies like ad hominem or strawman are seen as pattern‑of‑rhetoric issues rather than pure logic; they may sometimes carry relevant information (e.g., about credibility), complicating rigid classification.