Stop Clickbait: Detecting and Preventing Clickbaits in Online News Media

Abhijnan Chakraborty; Bhargavi Paranjape; Sourya Kakarla; Niloy Ganguly

In brief

Online news media outlets use sensationalized headlines (clickbaits) to increase reader clicks, exploiting the curiosity gap. This work identifies linguistic and structural features distinguishing clickbaits from legitimate headlines, builds an SVM classifier achieving 93% detection accuracy, and develops a Chrome extension that warns users and learns personalized blocking preferences achieving 89% accuracy.

Contributions

Linguistic and structural analysis of clickbait vs. non-clickbait headlines, revealing features like longer headlines, more stop words, hyperbolic language, and determiners in clickbaits
A classifier combining sentence structure, word patterns, clickbait language features, and n-gram features to detect clickbaits with 93% accuracy
Personalized blocking approaches using topical similarity and linguistic patterns to learn individual reader preferences
A browser extension that integrates the classifier and blocking mechanism across multiple websites

Method

The authors compiled 7,500 clickbait headlines from sites like BuzzFeed, Upworthy, and ViralNova, and 7,500 non-clickbait headlines from Wikinews. Linguistic analysis using Stanford CoreNLP revealed systematic differences: clickbaits average 10 words vs. 7 for non-clickbaits; use more contractions (22% vs. 0.6%), stop words (45% vs. 18%), and hyperbolic/positive-sentiment words; employ more determiners and pronouns; and feature longer syntactic dependencies.

The classifier uses 14 features across four categories: sentence structure (headline length, word length, stop-word ratio, dependency length), word patterns (contractions, punctuation, numbers), clickbait language (hyperbolic words, common phrases, internet slang, determiners), and n-gram features (word, POS, and syntactic n-grams scored by auxiliary Naive Bayes classifiers). An SVM with RBF kernel achieved best performance.

For blocking, the authors surveyed readers and found low agreement (Jaccard coefficient 0.13–0.15) on which clickbaits to block, motivating personalized classifiers. Two approaches extract topics from article headlines/metadata and HTML tags, expanding them via BabelNet; or normalize headlines to linguistic patterns and compute edit distance. A hybrid method combines both using equal weights.

Results

SVM achieved 93% accuracy, 0.95 precision, 0.90 recall in 10-fold cross-validation, significantly outperforming the rule-based Downworthy baseline (76% accuracy). In browser extension evaluation with 37 real users across Huffington Post, CNN, BuzzFeed, New York Times, and Facebook, the classifier maintained 94% accuracy. Among 16 users who explicitly blocked clickbaits, the personalized blocking component achieved 89% correctness according to user feedback.

Notes

Strength: Comprehensive linguistic analysis revealing systematic features of clickbait, strong empirical results, real-world deployment via Chrome extension with field evaluation. The personalization mechanism is well-motivated by the user study showing divergent preferences.

Limitations: Clickbait definition focuses on misleading sensationalism rather than outright falsehood; evaluated primarily on entertainment-oriented clickbait sites rather than political misinformation. The classifier is English-only, though the authors note the linguistic features are language-independent.