Guide
Everything you need to understand and use TellDear — from theoretical foundations to practical application.
The Architecture of Thought
Prefer a visual overview? Explore the 15-slide presentation →
1. What is TellDear?
TellDear is a Reasoning Taxonomy Explorer — an interactive platform for identifying, understanding, and analyzing reasoning flaws in text. It covers 535 distinct reasoning aspects across 6 dimensions, each with binary verification steps that can be evaluated by both humans and AI.
The platform combines insights from argumentation theory, cognitive psychology, propaganda studies, and formal logic into a single, searchable taxonomy. Whether you are a student learning about logical fallacies, a journalist fact-checking an article, or a researcher studying persuasion techniques, TellDear provides the tools to systematically analyze reasoning quality.
Key Features
- • Aspect Directory — Browse all 535 aspects, organized by dimension, with detailed descriptions and verification steps.
- • Text Analyzer — Paste any text and get an AI-powered deep analysis identifying specific reasoning flaws with quotes and explanations.
- • Knowledge Graph — Explore relationships between aspects, dimensions, and categories in an interactive force-directed graph.
- • Formal Logic (FOL) — Many aspects include First-Order Logic patterns and SMT verification for rigorous logical validation.
2. The Six Dimensions of Reasoning
The TellDear taxonomy organizes reasoning patterns into six dimensions. Each dimension captures a different category of reasoning flaws, from formal logic violations to subtle psychological biases.
D1: Logical Fallacies
(94 aspects)Violations of logical validity — arguments where the conclusion does not follow from the premises. Includes both formal fallacies (structural errors like affirming the consequent) and informal fallacies (content-based errors like ad hominem or straw man). These are the most classically studied reasoning errors, rooted in Aristotelian logic and modern Pragma-Dialectics.
D2: Manipulation & Propaganda
(96 aspects)Deliberate persuasion techniques designed to bypass rational evaluation. Includes emotional manipulation (fear appeals, loaded language), social pressure tactics (bandwagon, appeal to authority), and information distortion (cherry-picking, framing). These techniques are commonly found in political rhetoric, advertising, and media.
D3: Cognitive Biases
(131 aspects)Systematic deviations from rational judgment caused by the brain's mental shortcuts (heuristics). Unlike logical fallacies, cognitive biases are often unconscious and affect everyone. Includes anchoring bias, availability heuristic, confirmation bias, the Dunning-Kruger effect, and many more documented in cognitive psychology research.
D4: Statistical Errors
(131 aspects)Misuse or misinterpretation of data and statistical reasoning. Includes confusing correlation with causation, base rate neglect, survivorship bias, and cherry-picking data. These errors are particularly prevalent in science reporting, health claims, and economic arguments.
D5: Argumentation Schemes
(32 aspects)Patterns of reasoning that are legitimate in some contexts but fallacious in others. Argumentation schemes describe common inference patterns (argument from analogy, argument from consequences) along with critical questions that must be answered for the scheme to be valid.
D6: Discourse Mechanics
(51 aspects)Structural patterns in how arguments are constructed and presented in discourse. Includes framing effects, rhetorical questions used as assertions, goalpost shifting, and other conversational tactics that can undermine productive dialogue.
3. Understanding Aspects
An aspect is a single, well-defined reasoning pattern within the taxonomy. Each aspect has:
- Name & ID
- A human-readable name (e.g. "Ad Hominem") and a machine-readable identifier (e.g.
ad_hominem). - Dimension & Category
- Which of the six dimensions it belongs to, and its subcategory within that dimension.
- Verification Steps
- A sequence of binary (yes/no) questions that determine whether a given text exhibits this reasoning pattern. These form the core of the Atomic Instruction Dataset (AID).
- FOL Pattern
- A First-Order Logic formula representing the logical structure of the fallacy. For example, Affirming the Consequent:
(A ⇒ B) ∧ B ⇒ A. - SMT Expected
- The expected result from an SMT solver when verifying the FOL pattern — typically "invalid" for fallacies, indicating the reasoning is logically unsound.
The taxonomy currently contains 535 aspects with a total of 1838 verification steps. You can browse all aspects in the Aspect Directory.
4. How to Use the Analyzer
The Text Analyzer scans any text you provide against the full taxonomy. It operates in two modes:
Deep Analysis (AI)
When an API key is configured, the analyzer sends your text to an AI model along with the full taxonomy. The AI identifies reasoning flaws and returns:
- • An overall assessment of reasoning quality
- • Specific findings with exact quotes from your text
- • Explanations of why each finding is problematic
- • Severity ratings (high / medium / low)
- • Links to the corresponding aspect detail pages
Heuristic Mode (Fallback)
Without an API key, the analyzer uses keyword-based pattern matching. This mode is less accurate but works entirely client-side with no API calls.
- • Scans all 535 aspects using keyword heuristics
- • Shows detection confidence bars per aspect
- • Results grouped by dimension
- • Useful for quick, offline scanning
Step by Step
- Go to the Analyze page
- Paste your text into the input field (or click Example for a sample text)
- Click Deep Analyze (or Analyze in heuristic mode)
- Review the overall assessment for a summary of reasoning quality
- Examine each finding card: read the quoted passage, understand the explanation
- Click View aspect details to learn more about the specific reasoning pattern
5. Theoretical Background: Pragma-Dialectics
A key theoretical foundation of the TellDear taxonomy is Pragma-Dialectics, developed by Frans H. van Eemeren and Rob Grootendorst at the University of Amsterdam. This framework conceptualizes argumentation as a speech situation aimed at resolving a difference of opinion through rational discussion.
Pragma-Dialectics establishes ten prescriptive rules for critical engagement. Violations of these rules constitute logical fallacies — many of which are directly represented as aspects in the TellDear taxonomy.
The Freedom Rule
Parties must not prevent each other from advancing standpoints or casting doubt on them.
Violation: Ad Hominem, Straw Man
The Burden-of-Proof Rule
A party who advances a standpoint is obliged to defend it if requested.
Violation: Evading or Shifting the Burden of Proof
The Standpoint Rule
Attacks on a standpoint must relate to the actual standpoint advanced.
Violation: Straw Man
The Relevance Rule
Standpoints may only be defended using argumentation related to that standpoint.
Violation: Ignoratio Elenchi (irrelevant conclusion)
The Unexpressed Premise Rule
Parties may not falsely present something as an unexpressed premise or deny an implicit premise.
Violation: Denying an Implicit Premise
The Starting Point Rule
No party may falsely present a premise as an accepted starting point.
Violation: Arguing from unagreed-upon premises
The Argument Scheme Rule
A defense is only conclusive if it employs an appropriate, correctly applied argument scheme.
Violation: Faulty Analogy, Argumentum Ad Populum
The Validity Rule
Reasoning must be logically valid or capable of being made valid.
Violation: Hasty Generalization, confusing cause and effect
The Closure Rule
A failed defense must lead the protagonist to retract; a successful defense must lead the antagonist to retract doubt.
Violation: Refusal to Retract
The Usage Rule
Formulations must be clear and non-ambiguous.
Violation: Equivocation, Purposeful Ambiguity
6. Cognitive Biases & Heuristics
Cognitive biases are systematic deviations from rational judgment. Unlike logical fallacies (which are errors in argument structure), biases arise from the brain's mental shortcuts — heuristics that evolved to enable fast decision-making but can lead to predictable errors.
Information Processing
- Anchoring Bias
- Over-reliance on the first piece of information encountered.
- Availability Heuristic
- Overestimating the likelihood of events that come easily to mind.
- Apophenia
- Perceiving meaningful connections between unrelated phenomena.
- Functional Fixedness
- Limiting perception of an object to its traditional function.
Self-Perception
- Dunning-Kruger Effect
- Unskilled individuals overestimate ability; experts underestimate theirs.
- Bias Blind Spot
- Perceiving oneself as less biased than others.
- Overconfidence Effect
- Excessive confidence disproportionate to actual accuracy.
- Illusion of Transparency
- Overestimating how well others understand your mental state.
Social & Attribution
- Fundamental Attribution Error
- Overemphasizing personality-based explanations for others' behavior.
- Halo Effect
- A single positive trait influencing overall perception of character.
- Ingroup Bias
- Preferential treatment for perceived members of one's own group.
- Just-World Hypothesis
- Rationalizing injustices as deserved to maintain belief in a fair world.
Economic & Decision-Making
- Sunk Cost Fallacy
- Justifying continued investment based on cumulative prior investment.
- Loss Aversion
- Losses feel psychologically larger than equivalent gains.
- Hyperbolic Discounting
- Preference for immediate rewards over larger delayed ones.
- IKEA Effect
- Overvaluing products one has partially assembled.
7. Formal Logic: NL2FOL Pipeline
The NL2FOL (Natural Language to First-Order Logic) framework is a neurosymbolic pipeline that translates unstructured text into formal symbolic logic. It provides the formal backbone for verifying logical validity of arguments.
Pipeline Stages
SMT Verification
Logical validity is verified using the CVC4 SMT solver (Satisfiability Modulo Theory). The process works by checking the negation of the formula: if the negation is satisfiable, a counter-model is generated, identifying a logical fallacy.
; Example: Affirming the Consequent
; FOL: (A ⇒ B) ∧ B ⇒ A
; SMT checks: ¬((A ⇒ B) ∧ B ⇒ A)
; Result: SAT (satisfiable) → formula is INVALID → fallacy confirmed
Performance on Benchmarks
| Dataset | NL2FOL (GPT-4o) F1 | End-to-End LLM F1 |
|---|---|---|
| LOGIC | 78% | 96%* |
| LOGIC-CLIMATE | 80% | 58% |
* High end-to-end LLM score on LOGIC likely reflects training data leakage from public web sources.
8. Datasets & Corpora
LOGIC Dataset
2,449 examples of common logical fallacies across 13 categories (Ad Hominem, False Causality, False Dilemma, Faulty Generalization, Ad Populum, etc.).
LOGIC-CLIMATE
1,079 examples for out-of-domain generalization testing, using climate news metadata.
SNLI (Stanford NLI)
Provides "valid" (non-fallacious) benchmarks. The entailment class (~170,000 pairs) is used to construct valid reasoning examples.
COCOLOFA
News comments containing common logical fallacies, serving as a primary source for training on informal discourse.
9. Bibliography
Eemeren, F. H. van, & Grootendorst, R. (1996). Fundamentals of Argumentation Theory: A Handbook of Historical Backgrounds and Contemporary Developments. Lawrence Erlbaum Associates.
Eemeren, F. H. van, Grootendorst, R., & Henkemans, F. S. (2002). Argumentation: Analysis, Evaluation, Presentation. Lawrence Erlbaum Associates.
Iqbal, S., et al. (2023). Towards automated analysis of rhetorical categories in students essay writings using Bloom's taxonomy. In LAK 2023 Conference Proceedings (pp. 418-429). ACM. doi:10.1145/3576050.3576112
Lalwani, A., Kim, T., Chopra, L., Hahn, C., Jin, Z., & Sachan, M. (2024). Autoformalizing Natural Language to First-Order Logic: A Case Study in Logical Fallacy Detection. ACL Anthology. github.com/lovishchopra/NL2FOL
Hebb, D. O. (1949). The Organization of Behavior. Wiley & Sons.
Wikipedia. List of cognitive biases. en.wikipedia.org
10. Aspect Metadata & Tagging
Every aspect in the TellDear taxonomy carries three metadata tags that enable filtering, adaptive difficulty, and audience-appropriate recommendations. The tags are stored in the AID JSON and used throughout the platform.
Accessibility (1–3)
- 1Beginner — understandable without prior knowledge
- 2Intermediate — requires some critical thinking background
- 3Expert — requires academic background in logic, statistics, or psychology
Frequency (1–3)
- 1Common — appears daily in mainstream discourse
- 2Moderate — appears in structured argument and professional contexts
- 3Rare — specialized or academic, rarely in everyday communication
Subject Areas
1–4 applicable areas per aspect: politics, media, science, everyday, business, legal, education
Tags are generated by AI for 535 aspects and reviewed manually. They are stored in the AID JSON alongside definitions, verification steps, and FOL patterns — making them machine-readable for downstream applications.
AID JSON structure
{
"id": "ad_hominem",
"name": "Ad Hominem",
"tags": {
"accessibility": 1,
"frequency": 1,
"subject_areas": ["politics", "media", "everyday"]
},
...
}