Linguistic features¶
Linguistic features—including word choice, punctuation patterns, syntactic structure, readability metrics, and psycholinguistic properties—serve as signals for identifying deceptive or misleading content. These features are often more robust than content-based keywords and can generalize across domains.
Key papers¶
- Separating Facts from Fiction: Linguistic Models to Classify Suspicious and Trusted News Posts on Twitter — Combines linguistic features (bias markers, subjectivity cues, LIWC psycholinguistic features, moral foundation cues) with social network signals via neural networks for fake news classification. Shows linguistic features (hedges, factive verbs, assertive verbs) effectively distinguish verified vs. suspicious news; different suspicious types show distinct moral foundation appeals; syntax/grammar features degrade performance on noisy tweets.
- Automatic Detection of Fake News — Systematically analyzes linguistic differences between fake and legitimate news using LIWC psycholinguistic categories, readability metrics (Flesch-Kincaid, ARI), punctuation patterns, and syntactic features. Shows legitimate news uses more cognitive words and negations; fake news uses more positive emotions, certainty cues, and adverbs. Achieves 78% detection accuracy via readability features alone.
- Rashkin et al. (2017) — Truth of Varying Shades: Analyzing Language in Fake News and Political Fact-Checking: Linguistic analysis across satire, hoax, and propaganda categories revealing stylistic markers of unreliable news.
- A Stylometric Inquiry into Hyperpartisan and Fake News — Stylometric features (n-grams, readability indices) for differentiating hyperpartisan and fake news, showing style-based approaches have limited generalization.
Related topics¶
- Deception detection uses linguistic cues as core signals
- Fake news detection applies linguistic feature engineering at scale
- Stylometry focuses on authorship and writing-style analysis
Notes¶
Linguistic features have proven valuable because they capture the language use patterns of deceptive writers, who often employ distinctive patterns in emotional language, hedging, and structural properties. However, cross-domain generalization remains challenging because linguistic properties of fake content vary significantly across domains (e.g., sports fake news differs from political fake news in characteristic feature sets).