A WordNet-Aligned Taxonomy of Logical Fallacies for Low-Resource NLP

The article by Stanislav Penkov, “A WordNet-Aligned Taxonomy of Logical Fallacies for Low-Resource NLP,” proposes a compact and multilingual framework for identifying logical fallacies in natural language, with a particular focus on low-resource languages such as Bulgarian. The main motivation is that existing fallacy-detection datasets use inconsistent classification schemes and are predominantly English-language, making cross-dataset comparison and cross-lingual transfer difficult.

The study consolidates five existing datasets—Argotario, LOGIC_EDU, LOGIC_CLIMATE, MAFALDA, and CoCoLoFa—into a unified corpus. After normalization, the initial 12,531 examples are reduced to 11,111 usable instances belonging to a common nine-class target: eight logical fallacies plus a no fallacy background category. The eight fallacies are ad hominem, appeal to authority, appeal to emotion, appeal to popularity, false dilemma, hasty generalization, red herring, and slippery slope.

A central contribution is the introduction of a two-layer taxonomy. The first layer provides conceptual anchors by linking each fallacy category to concepts in WordNet 3.0. The second is an empirical lexical layer derived from the datasets themselves: distinctive words and bigrams for each fallacy are identified using log-odds weighting. For example, expressions related to expert testimony emerge for appeal to authority, vote majority for appeal to popularity, and easily lead for slippery-slope reasoning.

The WordNet alignment is especially important for cross-lingual transfer. Because WordNet synsets provide language-independent conceptual identifiers, they can be connected through Open Multilingual WordNet to lexical resources in other languages. The paper demonstrates this with Bulgarian seed terms—for example, concepts corresponding to majority, fear, expert, and consequence. These are intended as starting points for weak supervision, lexicon construction, annotation, retrieval, and prompting, rather than as automatically validated Bulgarian fallacy labels.

The study also highlights an important limitation of automated semantic alignment. A manual audit of 50 automatically generated cue-to-WordNet mappings found 52% acceptable, 22% borderline, and 26% incorrect. The author therefore emphasizes that the resource should function as a bootstrapping mechanism rather than a fully automatic fallacy classifier: human validation remains necessary, particularly when transferring concepts across languages.

Overall, the paper provides a semantic infrastructure for developing interpretable and multilingual NLP systems for detecting flawed or manipulative reasoning. Its relevance for Bulgarian lies in providing a systematic way to bootstrap resources even when large manually annotated Bulgarian datasets do not yet exist. At the same time, the paper explicitly notes that its Bulgarian examples demonstrate the projection mechanism and do not yet constitute either a validated Bulgarian fallacy dataset or a Bulgarian fallacy-detection classifier.

Suggested citation

Penkov, Stanislav. 2026. “A WordNet-Aligned Taxonomy of Logical Fallacies for Low-Resource NLP.” Proceedings of the Seventh International Conference on Computational Linguistics in Bulgaria (CLIB 2026).

Article: Full-text PDF

Proceedings: CLIB Conference Proceedings