PaperlayerView exact revision on arXiv

Source-linked AI summary

Automatically Identifying Fake News in Popular Twitter Threads

Cody Buntain, Jennifer Golbeck

arXiv:1705.01613v2cs.SI

TL;DR

Social-media scale makes manual assessment of inaccurate content difficult, motivating automated fake-news detection on Twitter. The paper transfers credibility models from CREDBANK and PHEME to BuzzFeed Twitter content, finding that crowdsourced-worker models outperform journalist-trained models and pooled models.

  • Problem

    Social-media volume makes it difficult for experts to manually assess and correct inaccurate content at scale.

  • Method

    The paper predicts Twitter-thread accuracy using credibility models trained on crowdsourced CREDBANK and journalist-labeled PHEME data, applied to BuzzFeed content.

  • Results

    Crowdsourced-worker models outperform journalist-trained and pooled models for classifying potentially fake Twitter threads.

  • Takeaways & Limitations

    Crowdsourced assessments can provide a useful, less expensive means to classify true and false Twitter stories rapidly and support users’ credibility judgments.

  • Takeaways & Limitations

    Structural differences between CREDBANK and PHEME may affect model transfer, making feature and cross-context differences attributable to dataset structure rather than model capability.

Abstract

from arXiv · show

Information quality in social media is an increasingly important issue, but web-scale data hinders experts' ability to assess and correct much of the inaccurate content, or `fake news,' present in these platforms. This paper develops a method for automating fake news detection on Twitter by learning to predict accuracy assessments in two credibility-focused Twitter datasets: CREDBANK, a crowdsourced dataset of accuracy assessments for events in Twitter, and PHEME, a dataset of potential rumors in Twitter and journalistic assessments of their accuracies. We apply this method to Twitter content sourced from BuzzFeed's fake news dataset and show models trained against crowdsourced workers outperform models based on journalists' assessment and models trained on a pooled dataset of both crowdsourced workers and journalists. All three datasets, aligned into a uniform format, are also publicly available. A feature analysis then identifies features that are most predictive for crowdsourced and journalistic accuracy assessments, results of which are consistent with prior work. We close with a discussion contrasting accuracy and credibility and why models of non-experts outperform models of journalists for fake news detection in Twitter.

I. INTRODUCTION

The paper addresses the challenge of assessing inaccurate social-media content at scale by automating fake-news detection on Twitter. It uses credibility assessments and analyzes differences between crowdsourced and journalist evaluations.

  • Social media volume has overwhelmed manual quality control, allowing questionable news to reach wide audiences.
  • The proposed classifier predicts whether Twitter conversation threads will be labeled accurate or inaccurate using credibility-inspired features.
  • The method transfers credibility models trained on PHEME and CREDBANK to BuzzFeed’s dataset of 35 highly shared true and false political stories.
  • The model correctly classifies two-thirds of Twitter fake-news stories and outperforms prior work.
  • Crowdsourced-worker models outperform journalist-trained models, while feature analysis finds different assessment patterns across groups.
  • The contributions include automated thread classification, feature analysis, and three aligned datasets capturing accuracy judgments.

II. RELEVANT WORK AND DATASETS

Prior work examined credibility through tweet, user, temporal, and propagation features, while this paper draws on PHEME’s journalist assessments and CREDBANK’s crowdsourced judgments. PHEME provides labeled rumor threads and annotations that quantify support, refutation, and disagreement.

  • Earlier studies evaluated tweet or user credibility and modeled rumor-propagation dynamics on social media.
  • Castillo et al. used 68 Twitter credibility features and achieved 64% accuracy with logistic regression on a selected feature set.
  • PHEME defines rumors as unverified, relevant statements that may later be confirmed true, false, or left unconfirmed.
  • PHEME contains 330 labeled source tweets across 140 stories, plus 4,512 descendant tweets organized into conversation trees.
  • Annotations record reply support, author confidence, and evidentiality, enabling disagreement and refutation to be quantified for accuracy modeling.
  • Of 330 PHEME conversation trees, 159 were true, 68 false, and 103 unverified.

B. The CREDBANK Dataset

CREDBANK is a large-scale Twitter dataset in which event-related topics receive crowdsourced accuracy assessments. Its ratings are highly skewed toward accuracy, requiring a majority-agreement threshold for credibility judgments.

  • CREDBANK covers 96 days, more than 1,000 event-related tweet sets, and approximately 37 million tweets assessed by 30 AMT annotators per event.
  • The construction filtered approximately one billion tweets, extracted 46,850 candidate event-topics, and identified 1,049 event-related topics through annotation.
  • CREDBANK provides a 30-element vector of accuracy labels rather than binary event-accuracy labels.
  • More than 95% of event annotations had a majority rating of “Certainly Accurate.”
  • Using a 70% majority-agreement threshold, 23% of events were classified as not perceived to be credible.

C. BuzzFeed News Fact-Checking Dataset

The paper frames fake-news detection as classifying popular Twitter stories by transferring credibility models to the BuzzFeed dataset. Its feature design spans structural, user, content, and temporal properties of Twitter activity.

  • BuzzFeed journalists collected over 2,000 posts from nine large verified Facebook pages and fact-checked them into four truth-status categories.
  • The central task is automatically classifying popular Twitter stories as accurate or inaccurate despite scarce labeled true and false stories.
  • The approach transfers credibility models trained on CREDBANK and PHEME to fake-news detection in BuzzFeed data.
  • The model uses 45 accuracy-prediction features grouped into structural, user, content, and temporal types.
  • Temporal features represent how structural, user, and content properties change over time, including tweet volume and author-age slopes.

1) Structural Features:

The paper represents Twitter conversations through structural, user, network, and content features that characterize thread form, participants, and expressed sentiment or disagreement.

  • Structural features: Structural features capture thread size, average tweet length, lifetime, conversation-tree depth, and frequencies or ratios of hashtags and media.These features are calculated across the entire conversation thread.
  • User features: User features describe account age, follower, friend, and authored-status counts, verification, and related account-time differences.
  • Network features: Network density represents users as graph nodes and mentions or retweets as edges, distinguishing dense internal interaction from sparser externally stimulated discussion.
  • Content features: Content features include tweet polarity, subjectivity, and disagreement measured by the amount of disagreement expressed in the conversation.

4) Temporal Features:

Temporal features describe how conversation characteristics change during a thread’s lifetime, while dataset alignment addresses differences in labels, sources, and tweet organization.

  • Temporal features: Temporal features accumulate content and activity values each minute, transform them logarithmically, and use linear-regression slopes to represent change over time.
  • Dataset alignment: The datasets require a consistent feature set and labels before their resulting models can be compared.
  • Dataset alignment: Because BuzzFeed data comes from Facebook while CREDBANK and PHEME are Twitter-based, the paper extracts matching Twitter threads rather than comparing the datasets directly.
  • Label alignment: CREDBANK accuracy assessments are converted into discrete labels because its Likert-scale annotator vectors differ from the class labels in PHEME and BuzzFeed.
  • Label alignment: CREDBANK labels use the bottom and top 15% of mean ratings: below 1.467 is negative and above 1.9 is positive.

3) Capturing Twitter’s Threaded Structure:

The paper converts heterogeneous tweet collections into comparable threaded conversations and infers disagreement annotations where datasets lack them.

  • Thread construction: PHEME threads begin with a popular root tweet and replies, whereas CREDBANK and BuzzFeed initially lack this threaded structure.
  • Thread construction: CREDBANK uses each event’s most-retweeted tweet as the root, and discards threads without reactions, leaving 115 positive and 95 negative samples.
  • Thread construction: BuzzFeed headline tweets serve as roots, with replies captured to mimic PHEME’s thread structure.
  • Disagreement inference: A disagreement classifier trained on PHEME support labels and IACv2 disputed labels supplies missing disagreement annotations for CREDBANK and BuzzFeed.
  • Disagreement inference: Stochastic gradient descent was the best tested disagreement predictor, achieving a mean 10-fold cross-validation ROC-AUC of 86.7%.Augmenting PHEME support labels with IACv2 was necessary to reach sufficient ROC-AUC of 72.66%.

C. Per-Set Feature Selection

The paper selects feature subsets separately for CREDBANK and PHEME, then evaluates their transfer to BuzzFeed using random-forest models and repeated ROC-AUC measurements.

  • Evaluation: Feature performance is measured with ROC-AUC using thirty instances of 10-fold cross-validation and a 100-tree random forest.ROC-AUC ranges from 0 to 1, with 0.5 representing a random coin toss for a balanced set.
  • Feature selection: Feature selection recursively removes the least performant feature using leave-one-out comparisons until one feature remains.
  • Feature selection: The selected feature subsets are constructed separately for CREDBANK and PHEME and may not overlap.
  • Transfer evaluation: Each source dataset trains a 100-tree random forest on its best feature subset, which is then applied to BuzzFeed using the corresponding features.
  • Transfer evaluation: The BuzzFeed evaluation repeats each training-and-application process 20 times and averages the resulting ROC-AUC values.

IV. RESULTS

Feature elimination identified compact, dataset-specific feature sets that predicted potentially false Twitter threads. When transferred to BuzzFeed stories, CREDBANK-based models performed best, while PHEME-only performance was substantially worse.

  • Feature Elimination: Seven of 45 features produced the highest-performing PHEME set, which achieved ROC-AUC 0.7407 and correctly identified 66.93% of potentially false threads.The features included media sharing, hashtag use, pronouns, disagreement, and the slope of authors’ friends over time.
  • Feature Elimination: Twelve of 45 features produced the most informative CREDBANK set, which achieved ROC-AUC 0.7184 and correctly identified 70.28% of potential false threads.The set included emoticons, mentions, punctuation, hashtags, media, account age, tweet length, followers, and verified authorship.
  • Feature Elimination: Only three features were shared by the crowdsourced-worker and journalist models: smile-emoticon frequency and media- or hashtag-tweet proportions.Both datasets’ top-ten features also contained five of Castillo et al.’s 16 best features.
  • Feature Elimination: The models’ 66.93% and 70.28% accuracies modestly exceeded Castillo et al.’s 61.81% accuracy, although the increases were marginal.The comparison is reported for PHEME and CREDBANK, respectively, against the prior model.
  • Adapting to Fake News Classification: On BuzzFeed, the CREDBANK-based model achieved ROC-AUC 73.80% and accuracy 65.29%, whereas the pooled model was near random and the PHEME model performed substantially worse.The pooled model scored ROC-AUC 53.14% and accuracy 51.00%; the PHEME model scored ROC-AUC 36.52% and accuracy 34.14%.

V. DISCUSSION

The discussion attributes the stronger crowdsourced-model performance to differences between perceived credibility and journalist-assessed factual accuracy. The aligned datasets support further study of these distinct assessment signals.

  • Contrasting Assessment Populations: Crowdsourced-worker models outperformed journalist-trained models when classifying popular Twitter stories as true or fake.The authors identify this as one of two significant discussion results.
  • Contrasting Assessment Populations: Limited feature overlap between PHEME and CREDBANK suggests that journalists and crowdsourced workers evaluate social-media accuracy differently.The discussion links this difference to the distinct assessment populations represented in the datasets.
  • Accuracy and Credibility: For headline-and-link threads, journalist-defined truth influences but does not dictate crowdsourced perceptions, which determine how stories are shared.The authors propose this as an explanation for CREDBANK models’ stronger BuzzFeed performance.
  • Accuracy and Credibility: Despite a bias toward believing what they read, CREDBANK assessors’ rating differences capture some latent feature of accuracy.The authors suggest future work could use this information to correct for non-experts’ bias.
  • Accuracy and Credibility: Journalists assess objective factual truth after consensus and review of event tweets, whereas CREDBANK workers assess rapid, perception-based believability.PHEME uses post-consensus journalistic judgments; CREDBANK emphasizes realtime responsiveness and perceived accuracy.
  • Implications: The aligned dataset enables future comparisons of assessment populations and may support education about signals used when evaluating social-media content.The discussion proposes informing both non-experts and journalists about the distinct features they use.

A. Limitations

The main limitations are structural differences between CREDBANK and PHEME and the focus on popular Twitter threads. The paper nevertheless frames popular fake stories as especially relevant because they have greater potential to misinform.

  • Dataset Alignment: Structural differences between CREDBANK and PHEME may confound feature-set and cross-context performance differences with dataset-generation effects.The authors propose evaluating the same rumors and fake-news threads with both assessor populations to avoid alignment biases.
  • Scope: The method applies only to popular Twitter conversation threads because it relies on highly retweeted threads and their features.Since most tweets are rarely retweeted, the method is usable on only a minority of Twitter conversations.
  • Scope: The popularity restriction is partly mitigated because popular fake stories have more potential to misinform than less popular fake stories.The paper states that non-retweeted fake news may lack traction or already be recognized as fake.
  • Conclusion: The authors present crowdsourced workers as a useful, less expensive means of rapidly classifying true and false Twitter stories.They suggest such systems could augment users’ credibility judgments.
Loading 1705.01613v2…