Source-linked AI summary
Stop Clickbait: Detecting and Preventing Clickbaits in Online News Media
Abhijnan Chakraborty, Bhargavi Paranjape, Sourya Kakarla, Niloy Ganguly
TL;DR
Online media rely on attention-generating headlines to attract clicks, while clickbaits can disappoint readers and distract attention. This paper develops automatic detection, personalized blocking, and a browser extension for warnings and blocking, achieving 93% detection accuracy and 89% blocking accuracy.
Problem
Clickbaits are widespread across websites, while fixed phrase and domain-based approaches are not scalable and may block non-clickbait articles.
Method
The paper extracts headline differences to detect clickbaits, builds personalized classifiers from readers’ histories, and integrates them into a browser extension.
Results
93% accuracy was achieved in detecting clickbaits and 89% accuracy in blocking them across offline and online experiments.
Takeaways & Limitations
The extension warns readers about clickbaits and automatically blocks similar headlines according to their stated preferences during future visits.
Takeaways & Limitations
The current classifier works with English headlines only, although the authors state its features can be extended to other languages.
Abstract
from arXiv · showhide
Most of the online news media outlets rely heavily on the revenues generated from the clicks made by their readers, and due to the presence of numerous such outlets, they need to compete with each other for reader attention. To attract the readers to click on an article and subsequently visit the media site, the outlets often come up with catchy headlines accompanying the article links, which lure the readers to click on the link. Such headlines are known as Clickbaits. While these baits may trick the readers into clicking, in the long run, clickbaits usually don't live up to the expectation of the readers, and leave them disappointed. In this work, we attempt to automatically detect clickbaits and then build a browser extension which warns the readers of different media sites about the possibility of being baited by such headlines. The extension also offers each reader an option to block clickbaits she doesn't want to see. Then, using such reader choices, the extension automatically blocks similar clickbaits during her future visits. We run extensive offline and online experiments across multiple media sites and find that the proposed clickbait detection and the personalized blocking approaches perform very well achieving 93% accuracy in detecting and 89% accuracy in blocking clickbaits.
1. Introduction
Online media use attention-attracting headlines to drive clicks, but clickbaits can disappoint readers and contribute to attention distraction. The paper proposes a web-wide system combining clickbait detection with personalized blocking and evaluates it across multiple media sites.
- Online outlets use catchy headlines to compete for reader attention and attract article clicks that support advertising revenue.
- Clickbaits exploit the Curiosity Gap by using forward references that compel readers to click to fill a knowledge gap.
- Clickbaits may disappoint readers and contribute to attention distraction, cognitive overload, and reduced attention to in-depth news.
- Existing fixed phrase and domain rules are difficult to scale and can block non-clickbait articles, motivating a comprehensive web-wide solution.
- The proposed system combines automatic headline detection, reader-specific blocking predictions, and a browser extension that warns about and blocks clickbaits.
- 93% accuracy was achieved for clickbait detection and 89% accuracy for personalized clickbait blocking across offline and online experiments.
2. Dataset
The dataset includes clickbait and non-clickbait headlines, with non-clickbait headlines drawn from 18,513 Wikinews articles and clickbait labels produced by volunteer annotation. The classifier comparison used balanced samples of 7,500 articles from each category.
- 18,513 Wikinews articles supplied the non-clickbait headlines for the dataset.
- Six volunteers labeled headlines from selected domains as clickbait or non-clickbait, with each article labeled by at least three volunteers.
- Fleiss’ κ was 0.79, indicating substantial inter-annotator agreement for the headline labels.
- 7,623 articles were marked as clickbaits using the majority vote as ground truth.
- 7,500 articles from each category were randomly selected to create equal clickbait and non-clickbait representations for comparison and classifier construction.
3. Comparing Clickbaits and Non-Clickbaits
Clickbait headlines differ from conventional news headlines in length, wording, syntax, subject choice, and grammatical patterns. They more often use function words, repeated phrases, complex dependencies, and reader-oriented references, while non-clickbait headlines emphasize concise factual content.
- Clickbait headlines average 10 words, compared with 7 for non-clickbait headlines.
- 22% of clickbait headlines contain word shortenings, compared with 0.6% of traditional news headlines.Clickbait headlines also have shorter average words: 4.5 characters versus 6.
- Clickbaits have longer syntactic dependencies than non-clickbaits, reflecting more complex phrasal sentences.Their average syntactic parse-tree height is 10.03, compared with 6.45 for non-clickbaits.
- 45% of words in clickbait headlines are stop words, compared with 18% in non-clickbait headlines.Because stop words contribute to clickbait sentence semantics, they were retained in subsequent n-gram analysis.
- 62% of clickbait headlines contain one of the 40 most common clickbait subject words, compared with 16% of non-clickbait headlines.Clickbaits also use more determiners, personal and possessive pronouns, adverbs, and verbs, whereas non-clickbaits contain more proper nouns.
- The top 0.05% of unique n-grams occur in 65% of clickbait headlines, compared with 19% of non-clickbait headlines.This repetition contrasts with the more factual and unique phrasing of conventional news headlines.
4. Classifying Headlines as Clickbaits
The paper classifies headlines using linguistic and structural differences between clickbait and non-clickbait, achieving strong performance against a fixed-rule baseline.
- Feature Design: The classifier uses structural, word-level, language, and n-gram features to distinguish clickbait from non-clickbait headlines.Features include headline and word lengths, stop-word ratios, syntactic dependencies, punctuation, contractions, hyperbolic words, clickbait phrases, slangs, determiners, and pruned n-grams.
- Evaluation: Three prediction models are evaluated with 10-fold cross validation using individual feature categories and their combination.The models are SVM with an RBF kernel, Decision Trees, and Random Forests, evaluated using Accuracy, Precision, Recall, F1, and ROC AUC.
- Results: 76% accuracy, 0.72 recall, and 0.78 precision are achieved by the fixed-rule Downworthy baseline.The proposed classification technique outperforms this baseline by a large margin.
- Scope: The current classifier operates on English headlines, although the authors state its linguistic features can extend to other languages.The stated cross-language extension is based on the authors’ view that the relevant characteristics are common linguistic phenomena.
5. Blocking Clickbait Headlines
The paper models each reader’s clickbait-blocking preferences using topical similarity, linguistic patterns, or a hybrid of both. The pattern-based approach performs better and faster, making it suitable for real-time blocking.
- Personalization: Reader preferences are personalized because the survey found substantial variation in which clickbait headlines readers wanted to block.Twelve readers reviewed 200 randomly selected clickbait headlines and marked those they would click or block.
- Personalization: The blocking task predicts whether a reader wants to block a new clickbait based on articles previously clicked or blocked.The approach models reader interests from past behavior and considers both topical and linguistic interpretations of preferences.
- Blocking Approaches: The topical approach compares a new article’s topics with topics extracted from the reader’s clicked and blocked articles.Topics come from headline content words, article metatags, and HTML keywords; BabelNet expands them into merged semantic nuggets.
- Blocking Approaches: The pattern-based approach compares normalized headline patterns using word-level edit distance and then makes the block or do-not-block decision.Patterns are formed separately from blocked and clicked articles, using the latest top 100 links ordered by timestamps for tag extraction and nugget formation.
- Results: The pattern-based approach yields better results and executes faster than the more involved topic-based approach.Because it is better suited to real-time environments, it was integrated into the Stop Clickbait browser extension.
6. Browser Extension: Stop Clickbait
Stop Clickbait is a Chrome extension that detects clickbait links across webpages, warns users, and learns individualized blocking choices for future visits. In online use across multiple sites, it achieved strong classification and blocking performance.
- Extension Functionality: The extension warns users about clickbaits on different webpages and automatically blocks similar clickbaits after users choose to block them.It is intended to increase applicability across websites by combining the classifier with personalized blocking.
- Extension Functionality: The extension scans webpage and dynamically inserted anchor elements, sending anchor text or URLs to a server for classification.The classifier predicts from anchor text when available and otherwise uses the URL-derived webpage title.
- Extension Functionality: Classified clickbait links receive green indicators, which expose options to block similar content or report a misclassification.Links not classified as clickbait remain unmarked.
- Personalization: Each extension instance receives a unique 32-byte identification number to support training for its personalized classifier.
- Online Evaluation: 94% accuracy and 0.934 F1-score are achieved across all websites visited by users.The online evaluation involved 37 users during April 2016 and covered multiple domains, including Facebook, Buzzfeed, and the New York Times.
- Online Evaluation: 89% of links are correctly blocked on average according to user feedback.Sixteen of the 37 users explicitly blocked more than 10 clickbaits, after which personalized classifiers were used during further visits.
7. Related Work
Prior work addressed clickbait detection or removal through platform-specific systems and fixed rules, but the paper presents a broader cross-web solution. The related literature also connects clickbaits to tabloidization, forward referencing, and misleading content.
- Background: Media-studies research links clickbaits to tabloid properties such as simplification, spectacularization, sensationalization, and reduced depth of reporting.
- Background: Psychological research describes clickbaits as using discourse deixis and cataphora for forward referencing, while another study frames them as misleading content or false news.
- Existing Detection: Facebook and prior Twitter research attempted clickbait detection using behavioral ratios, time spent, common words, and tweet-specific features.
- Existing Systems: Downworthy detects headlines with fixed clickbait phrases, while other approaches block links from fixed domains or operate on a single domain.These approaches may miss website-specific nuances, block non-clickbait links, or require continual rule tuning.
- Positioning: The paper proposes a comprehensive solution designed to work across the web rather than relying on fixed rules or domain lists.
8. Conclusion
The paper combines clickbait detection with personalized blocking in a Chrome extension that warns readers and blocks similar clickbaits on later visits. It presents the work as an initial comprehensive solution and releases data and source code to support further development.
- The Chrome extension warns readers about clickbaits across different media websites and lets them block unwanted clickbaits.
- The paper presents its approach as the first comprehensive attempt to deter the prevalence of clickbaits.
- The authors identify improving classification and blocking performance and tuning the extension with further user feedback as future work.
- The data and source code are publicly available to encourage researchers and developers to contribute to combating clickbaits.