Satire detection¶

Satire detection is the task of identifying satirical or parodical news articles—pieces written as humorous critique that may be misinterpreted by readers as legitimate news. Unlike outright falsehoods, satire relies on absurdity, irony, and exaggeration; the challenge is that readers may not recognize the mocking intent, leading to unintended spread as misinformation. Survey evidence shows up to 28% of Republican respondents and 14% of Democratic respondents reported believing stories from the Babylon Bee (a prominent satire site) as "definitely true."

Methods¶

Text-based approaches exploit linguistic and semantic patterns: - Classical ML: SVM with bag-of-words, n-grams, lexical features (Burfoot & Baldwin 2009; Rubin et al. 2016) - Deep learning: Hierarchical RNNs with attention mechanisms and handcrafted linguistic features (Yang et al. 2017)

Multi-modal approaches recognize that satirical news articles often pair absurd or manipulated images with headlines: - ViLBERT-based fusion (Li et al. 2020) achieves 93.8% accuracy by jointly modeling images and text via early, deep cross-modal attention - Image forensics alone (Error Level Analysis + CNN) underperforms, suggesting that satire requires joint reasoning across modalities

Key papers¶

[[2019-levi-fake-news-satire-nuances]]: Text-based satire vs. fake news distinction; proposes semantic and linguistic cues.
A Multi-Modal Method for Satire Detection using Textual and Visual Cues: Multi-modal satire detection using ViLBERT; demonstrates that visual and textual fusion outperforms uni-modal baselines.
Burfoot & Baldwin (2009) — early SVM-based approach with handcrafted features.
Yang et al. (2017) — hierarchical attention mechanism for satire.

Connections¶

Misinformation and fake news detection — satire is a distinct phenomenon; misinterpreted satire becomes misinformation.
Multimodal fake news detection — images in satire articles convey absurdity.
Humor Detection — satire relies on irony and humor; related computational methods.
Content-based fake news detection — text and image analysis without relying on source labels.

Notes¶

Satire detection has received less research attention than fake news detection, partly because major news platforms address satire through source labeling (tagging articles from known satire outlets). However, automated detection remains valuable when sources are unknown or when satire spreads across platforms lacking source context. The multi-modal nature of satire—where image manipulation is deliberate and often essential to the joke—makes this a natural domain for vision-language models.