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

Abstract

We present a compact taxonomy of eight logical fallacies designed for the analysis of manipulative reasoning and argumentative discourse, with a specific focus on low-resource settings. The taxonomy consolidates multiple public datasets into a nine-class training target (eight fallacies plus a background label no fallacy), with class distributions documented to support future stratified sampling or class-weighted modeling, and organizes these classes within a small parent–child hierarchy (relevance, presumption, causality, reasoning form) used for interpretability only (not as training labels), semantically grounding the target categories in WordNet 3.0 via synset IDs and offsets. From corpus evidence, we extract representative lexical cues using log-odds weighting with lemmatization and bigram detection, and link them to synsets to form a two-layer conceptual–lexical mapping. Crucially, the WordNet anchoring enables cross-lingual projection via Open Multilingual WordNet (OMW), supporting the creation of candidate seed lexicons and potential weak-supervision signals without prior in-language fallacy annotation. The framework thus provides a practical recipe for taxonomy-based lexical-semantic resource development, multilingual bootstrapping, and interpretable, fallacy-aware natural language processing (NLP).