Skip to content

Separating Facts from Fiction: Linguistic Models to Classify Suspicious and Trusted News Posts on Twitter

Authors: Svitlana Volkova, Kyle Shaffer, Jin Yea Jang, Nathan Hodas

Venue: ACL 2017 (Short Papers), pages 647–653, Vancouver, Canada, July 30 – August 4, 2017 — DOI

Organization: Pacific Northwest National Laboratory, Data Sciences and Analytics Group

TL;DR

This paper presents neural network models combining tweet content, social network interactions, and linguistic features (bias markers, subjectivity cues, moral foundations) to classify Twitter news posts as suspicious (propaganda, hoaxes, clickbait, satire) or verified. The models achieve 95% accuracy on binary classification and 0.7 F1-macro on multi-class classification. Linguistic analysis reveals that verified news uses significantly less biased, subjective language and appeals to different moral foundations than suspicious news types.

Contributions

  • Dataset: Manually curated labeled Twitter dataset of 130,000 retweets from 174 suspicious news accounts (99 propaganda, 9 satire, 8 hoaxes, 18 clickbait) and 252 verified news accounts, collected during Brussels bombing crisis (Mar 15-29, 2016).
  • Architectures: LSTM and CNN-based neural networks with late fusion technique combining three feature streams: tweet embeddings, social graph signals (mentions), and linguistic cues.
  • Baselines comparison: Neural networks outperform logistic regression baselines by 0.2 accuracy on binary task and 0.07 F1-macro on multi-class task.
  • Feature analysis: Social interaction features prove most informative for distinguishing suspicious news types; linguistic features improve all models; syntax/grammar features (unlike prior deception detection work) degrade performance on tweets.
  • Linguistic analysis: Quantitative statistical analysis of moral foundations, bias markers (hedges, assertive verbs, factive verbs, implicative verbs), and subjective language across all news types.

Method

Data collection: 130K tweets from Twitter firehose (Mar 15–29, 2016) for 174 suspicious and 252 verified news accounts. Labels propagated from account classification to retweets. Standard NLP preprocessing, part-of-speech tags, and dependency parses extracted via SyntaxNet.

Features: - Content: GloVe embeddings (200-dim) on tokenized tweet text - Social graph: One-hot vectors of user interactions (mentions), excluding source account mentions to prevent overfitting - Bias cues: Hedges (expressions of tentativeness), assertive verbs (certainty levels), factive verbs (presupposed truth), implicative verbs (truth implications), report verbs - Subjectivity cues: Strongly and weakly subjective words from opinion lexicons, positive/negative opinion words - Psycholinguistic cues: LIWC features (statistics, rhetorical questions, imperatives, personal pronouns, emotional language, quotations, certainty markers) - Moral foundation cues: Five moral value dimensions (care/harm, fairness/cheating, loyalty/betrayal, authority/subversion, purity/degradation) hypothesized to differ between suspicious and verified news

Neural architecture: Two-stream fusion model with separate content subnetwork (embedding → LSTM or 1D convolution + max pooling) and network subnetwork (feed-forward on interaction vectors), concatenated with linguistic/moral feature vectors into dense layers (100 units) with sigmoid/softmax output activation. Trained 10 epochs with ADAM optimizer, 10-fold cross-validation.

Results

Classification (Table 2): - Binary task (suspicious vs. verified): 0.95 accuracy with RNN + network + linguistic features (best); 0.98 AP, 0.99 F1 - Multi-class task (4 suspicious types): 0.7 F1-macro with RNN + network alone (best); 0.91 accuracy - Neural models significantly outperform logistic regression on Doc2Vec or TFIDF features - Adding network features improves both tasks (~0.05 accuracy binary, ~0.02 F1 multi-class) - Adding linguistic cues boosts performance across all models - Negative result: Syntax/grammar features decrease performance by 0.02 accuracy (binary) and 0.02 F1 (multi-class), attributed to tweets being shorter, noisier, and harder to parse than product reviews (prior work domain)

Linguistic analysis (Table 3): - Verified vs. suspicious: Verified news contains significantly fewer bias markers (hedges, assertive verbs), less subjective language, and less harm/care, loyalty/betrayal, authority moral cues - Propaganda: Targets moral foundations more than satire/hoaxes; appeals to authority and fairness more than satire - Satire: Most linguistically distinct from propaganda/hoaxes; uses more loyalty/betrayal moral appeals and more subjective language - Hoaxes: Fewest bias markers compared to satire; similar to propaganda on moral and subjectivity cues - Clickbait: Targets moral foundations least; uses most subjective cues among suspicious types

Retweet patterns (Figure 3): Propaganda, clickbaits, and hoaxes show high-volume spikes over short periods; rumors remain active over longer timeframes; propaganda and rumors focus on Brussels bombing topics while clickbaits/hoaxes cover divergent topics.

Connections

Notes

This paper makes strong methodological and analytical contributions by systematically combining content, network, and linguistic features for fine-grained suspicious news classification. The finding that social interaction signals are most informative for distinguishing suspicious news types (rather than just verified vs. suspicious) is valuable and suggests that propagation patterns encode distinct deception strategies.

The linguistic analysis is particularly novel and rigorous—the paper goes beyond classification performance to characterize how propaganda, satire, hoaxes, and clickbait differ in their appeal to moral values and use of persuasive language. The negative result on syntax/grammar features is also instructive: deception detection strategies that work on formal prose (product reviews) may not transfer to the noisy, constrained space of tweets.

Limitations include the account-level rather than tweet-level annotation (propagating labels from account to all retweets introduces noise) and the single crisis period (Mar 2016), which may limit generalizability. The paper notes that suspicious news annotations should ideally be at tweet level but acknowledges the cost. The use of established Twitter accounts for ground truth (PropOrNot for propaganda, fakenewswatch for others) is reasonable but manual verification of labels strengthens the contribution.