Cross-SEAN: A Cross-Stitch Semi-Supervised Neural Attention Model for COVID-19 Fake News Detection

1. Introduction

• The paper defines COVID-19 fake tweets as tweets contradicting statements released by governmental health organisations.Genuine tweets are taken from those organisations’ official accounts.Source: §1 · pp. 2, 5
• COVID-19 misinformation is presented as a public-health concern that can spread through social media and put lives at risk.The introduction notes that uninformed sharing and the lack of verification barriers contribute to misinformation circulation.Source: §1 · p. 2
• CTF combines 45.26K labeled tweets—18.55K genuine and 26.71K fake—with 21.85M unlabeled tweets.The four-stage collection process uses health-organization statements, BERT and RoBERTa to identify supporting or contradicting claims, and manual annotation.Source: §1 · p. 5
• Cross-SEAN combines tweet text, user and tweet metadata, external stances, and unlabeled data in a cross-stitch semi-supervised attention model.External knowledge from trusted domains may include information absent from training data, partially helping early detection of emerging fake news.Source: §1 · p. 5
• 0.95 F1 Score on CTF; Cross-SEAN outperforms seven baselines by at least 9%.The paper compares it with seven state-of-the-art fake-news detection models.Source: §1 · p. 5
• Chrome-SEAN is a Chrome extension that classifies tweets on their Twitter pages and gathers human feedback for confidence-conditioned online learning.The extension is deployed to handle concurrent requests.Source: §1 · p. 6

2. Related Work

• Earlier fake-news studies use linguistic, temporal, user, structural, visual, and multimodal features, with methods ranging from recurrent and convolutional models to graph-based approaches.Other lines of work address satire and humor, n-gram analysis, and combinations of text mining and supervised algorithms [28].Source: §2 · pp. 6–7
• Semi-supervised text-classification work is grouped into multi-view, data-augmentation, and transfer-learning approaches.Examples include hidden-space Mixup augmentation [50] and ULMFiT transfer learning [54].Source: §2 · pp. 7–8
• The authors identify limited labeled data, unexplored combinations of multiple features, and no published COVID-19 fake-news detection work as gaps.They address these gaps with CTF, unlabeled-data use, and cross-stitch input combination.Source: §2 · p. 8

Dataset collection and annotation

• Stage 1 combined public COVID-19 tweet datasets, tweet IDs, hashtag-based Twitter API collection, and tweets from official health organizations.Tweets from the official accounts were treated as genuine.Source: §3 · pp. 8–9
• Stage 2 gathered labeled COVID-19 claims and URLs from fact-checking sites, supplemented by official health-body material.These supporting statements and URLs were used in the subsequent labeling stage.Source: §3 · p. 9
• Stage 3 labeled URL-sharing tweets according to the URL class, yielding 5.3K fake and 7.5K genuine tweets, and added 10K genuine organization tweets.The authors note that assigning a tweet the shared URL’s class could introduce noise, but manual inspection found the assumption held for most cases.Source: §3 · p. 9
• BERT similarity labeling added 9.7K fake tweets, while RoBERTa sentence-pair inference added 10.6K fake tweets.BERT used a cosine-distance threshold of 0.9; RoBERTa, fine-tuned on SNLI, used contradiction and entailment results.Source: §3 · p. 9
• Three expert annotators manually verified 16,000 of 45,261 labeled tweets, with 8,000 each labeled fake and genuine and inter-annotator agreement of 0.82 Krippendorff’s α.When tweet text alone was insufficient for confident annotation, annotators could consult tweet and user features.Source: §3 · pp. 9–10
• Automated Stage 3 labels matched human annotations on an average of 92% of 16,000 samples, indicating 8% noise in the automated labels.The authors describe the pipeline as fully automated and fast, enabling a relatively large labeled corpus.Source: §3 · p. 10

Presence of hashtags

• Fake tweets more often featured #china and #bioweapon, while #WHO was more prominent in genuine tweets; #trump appeared much more in fake tweets.The authors note recurring #vaccine use in fake tweets and dominance of #togetherathome, #stayhome, and #socialdistancing in genuine tweets.Source: §4 · p. 11

Presence of URLs

• 14,830 genuine and 8,761 fake tweets contained URLs, averaging 0.87 and 0.35 URLs per tweet, respectively.The authors suggest the contrast may indicate a greater tendency for genuine tweets to support claims.Source: §4 · p. 11

Presence of multimedia and sentiment

• Fake tweets averaged 0.05 sentiment polarity versus 0.096 for genuine tweets; fake tweets were more grouped in highly negative and neutral sentiment zones.The polarity scale was -2 to 2; fake tweets also contained 2,491 media files across 2,344 tweets, versus 1,473 media files in genuine tweets.Source: §4 · p. 11

Likes and retweets

• 142.65 average likes per genuine tweet compared with 4.25 per fake tweet; 64% of fake tweets and 35% of genuine tweets were retweets.The authors attribute the higher genuine-tweet average to the many tweets from popular public health organizations.Source: §4 · p. 12

Visual representations

• Labeled text representations showed substantial genuine/fake overlap, whereas unlabelled text representations showed topic polarization; fake-tweet users clustered in the labeled user-feature space.The figures visualize tweet text, tweet features, and user features for labeled and unlabelled data.Source: Figure 4 · p. 11, §4 · p. 12, Figure 5 · p. 12

5.1. Explicit Tweet and User Features and External Knowledge

• The model combines 10 handcrafted tweet features and 8 user features with tweet text, and adds up to 10 retrieved context sentences selected by BERT-embedding similarity.Source: §5.1 · p. 13
• Cross-SEAN uses one half of the unlabelled data to fine-tune word embeddings and the other half for unsupervised adversarial training.Source: §5.1 · p. 13

5.2. Model Architecture

• Separate tweet- and user-feature streams pass through cross-stitch units, while transformed text, external-knowledge, and user-feature representations are concatenated for classification.Source: §5.2 · pp. 15–16

5.3. Training Strategies

• Training combines maximum-likelihood loss with adversarial training on labelled data and virtual adversarial training to add robustness against adversarial inputs.Source: §5.3 · pp. 16–17

• Table 2 presents competing models’ features and performance comparison on CTF, with TT meaning Tweet Text, TF Tweet Features, UF User Features, and UL Unlabelled Data.Source: Table 2 · p. 17

6. Experimental Setup and Results

• Experiments use a single-layer Bi-LSTM with maximum sequence length 128, hidden dimension 512, 300-dimensional embeddings, and Adam at learning rate 0.001.Source: §6 · p. 17

6.1. Comparative Evaluation

• Cross-SEAN beats all baselines by a margin of at least more than 6% accuracy and 9% F1 Score, with dEFEND the best baseline.Source: §6.1 · p. 18

6.2. Ablation Study

• A steady performance increase accompanies adding semi-supervised losses to the supervised maximum-likelihood objective; adversarial losses also change loss-curve behavior.Source: §6.2 · pp. 18–19, Figure 8 · p. 19
• The best reported architecture uses cross-stitching between tweet and user features, combining one output early and the other later; attention also improves final-model performance.Source: Table 4 · p. 18, §6.2 · p. 19

Purpose

• Chrome-SEAN deploys Cross-SEAN for real-time identification of fake tweets through a browser extension.The extension is intended to give users easy access to warnings while using Twitter.Source: §7 · p. 20

Extension implementation

• The extension sends tweet IDs to a Flask-hosted Cross-SEAN API, with Redis handling load balancing across concurrent requests.jQuery handles POST API requests between the extension and server.Source: §7 · p. 20

Extension workflow

• Chrome-SEAN scans a tweet ID from the URL or accepts one manually, then displays the model’s predicted class and confidence.Browser-side tweet extraction is instant; verification averages 1.2 seconds per single API request.Source: §7 · p. 20

Feedback and online training

• User feedback is added to the extended online dataset, and the model trains on differing feedback only when its confidence is below 0.6.The authors describe safeguards against attackers attempting to pollute results.Source: §7 · pp. 20–21

User Study

• 203 of 215 ratings were positive, corresponding to 94% accuracy and an F1 Score of 94.3%.The study involved 35 users testing 215 tweets.Source: §7 · p. 21

Dataset contribution

• CTF is presented as the first labeled Twitter dataset of COVID-19 genuine and fake tweets, with unlabelled data and analysis of surface-level linguistic features.The tweets were collected from diverse sources and human-annotated.Source: §8 · p. 21

Model

• Cross-SEAN combines tweet text, tweet and user features, and relevant external knowledge in a semi-supervised cross-stitch attention model.It leverages labelled and unlabelled data, with supervised maximum-likelihood and adversarial losses and virtual adversarial training for the unsupervised loss.Source: §8 · p. 21

Results

• Cross-SEAN is described as highly effective and outperforming seven state-of-the-art models; the authors also report ablations of network combinations and objective-function losses.The ablations examine component importance and variation across mixed-objective configurations.Source: §8 · pp. 21–22

Deployment

• Chrome-SEAN applies Cross-SEAN in a browser extension for flagging fake tweets, collecting user feedback, and online training when model confidence is low.The authors report reasonable performance in a small-scale user study.Source: §8 · p. 22

Limitations

• Twitter’s microblog language can make external knowledge noisy, and political bias in some trusted online sources can introduce bias into that knowledge.Source: §8 · p. 22
• External knowledge may help with emerging fake news but does not promise complete robustness or early detection.Source: §8 · p. 22
• Cross-SEAN is tuned specifically for COVID-19 fake news and was not tested on general fake news on Twitter.Source: §8 · p. 22

Future work

• Future work includes modeling follower-followee and retweet networks for early detection, filtering external knowledge for bias and noise, improving explainability, and using media such as images, GIFs, and videos.Source: §8 · p. 22