False News Detection on Social Media

Juan Cao, Qiang Sheng, Peng Qi, Lei Zhong, Yanyan Wang, Xueyao ZhangarXiv 2019 · arXiv

TL;DR

This research introduces MCG-FNeWS, a large-scale multi-modal dataset and competition framework designed for real-time false news detection on social media. By integrating text, images, publisher profiles, and external debunking knowledge, the study demonstrates that multi-modal fusion is the most effective approach for identifying misinformation before it goes viral.

The Challenge of Real-Time Detection

False news spreads nearly six times faster than factual information on social media platforms. Traditional detection methods often rely on "social context"—analyzing user replies, retweets, and propagation patterns over time. However, by the time sufficient social data is collected to flag a post, the misinformation has often already reached a massive audience and caused significant damage.

The authors propose a shift toward real-time detection based on the raw content of a post at the moment of publication. This is increasingly difficult as false news becomes more sophisticated, utilizing tampered visuals and mimicking legitimate journalistic styles to evade simple text-based filters.

Methodology and Subtasks

The study evaluates the contribution of different data types through three distinct subtasks:

  1. Subtask A (Textual Analysis): Identifying writing styles and linguistic features characteristic of misinformation.
  2. Subtask B (Visual Analysis): Detecting misrepresented or manipulated images.
  3. Subtask C (Multi-modal Fusion): Combining text, images, and publisher profiles (user metadata) for a comprehensive determination.

To establish a benchmark, several baseline architectures were tested:

  • Text Models: LSTM, GRU, TextCNN, and BERT.
  • Image Models: VGG19 and the Multi-view Visual Neural Network (MVNN).
  • Fusion Strategies: Early fusion (feature concatenation), Late fusion (prediction averaging), and an attention-based Recurrent Neural Network (attRNN).

A critical component of this framework is a debunking repository containing over 37,000 refutations, which encourages models to leverage external factual knowledge rather than relying solely on the post's internal content.

The MCG-FNeWS Dataset

The researchers introduced MCG-FNeWS, a dataset curated from Sina Weibo (2012–2018), representing one of the largest multi-modal resources in the field.

  • Training Set: 38,471 posts with a balanced split between real and false news.
  • Validation/Testing: Approximately 8,000 additional posts.
  • Hard Negatives: To ensure robustness, the dataset includes "real news" samples that share linguistic similarities with false news, preventing models from relying on simple clickbait indicators.

Key Insights

  • The Superiority of Fusion: Combining multiple modalities consistently improves performance. While text remains a strong signal, adding visual data and publisher metadata provides a more holistic view of the post's credibility.
  • Early Fusion Efficacy: In baseline tests, "Early Fusion" (concatenating text and image features before classification) surprisingly outperformed more complex attention-based models like attRNN.
  • Visuals as a Weak Link: Image-only detection (Subtask B) performed significantly worse than text-only methods. This suggests that while images are used to attract attention, the deceptive signal is often more pronounced in the text or the incongruity between the text and the image.
  • The Importance of Publisher Context: Metadata regarding the account posting the news adds a vital layer of context that content-only models might miss.

Experimental Results

The baseline leaderboard highlights the performance of various models across different modalities, confirming that multi-modal approaches yield the best results.

Performance comparison of text, image, and multi-modal fusion methods on the MCG-FNeWS dataset.

Key findings from the results include:

  • Text-only: BERT emerged as the top performer with 86.7% accuracy.
  • Image-only: MVNN achieved 80.5% accuracy, significantly outperforming standard VGG19 architectures.
  • Multi-modal: Early Fusion reached the highest overall accuracy at 87.6%.
  • Precision vs. Recall: TextCNN achieved the highest precision (95.3%), indicating high reliability when flagging news as fake, though it had lower recall compared to BERT or LSTM.

Limitations and Future Directions

Despite the dataset's scale, several challenges remain:

  • Knowledge Gaps: Not all false news has a corresponding entry in the debunking repository, limiting the effectiveness of models that rely heavily on external fact-checking.
  • Cross-Platform Generalization: As the data is sourced from Sina Weibo, tactics and styles may vary on other platforms like Twitter or Instagram.
  • Explainability: While the competition encourages providing explanations for flags, quantitative evaluation currently focuses on accuracy, leaving qualitative interpretability as a secondary goal for future research.

Summary updated

Abstract

Social media has become a major information platform where people consume and share news. However, it has also enabled the wide dissemination of false news, i.e., news posts published on social media that are verifiably false, causing significant negative effects on society. In order to help prevent further propagation of false news on social media, we set up this competition to motivate the development of automated real-time false news detection approaches. Specifically, this competition includes three sub-tasks: false-news text detection, false-news image detection and false-news multi-modal detetcion, which aims to motivate participants to further explore the efficiency of multiple modalities in detecting false news and reasonable fusion approaches of multi-modal contents. To better support this competition, we also construct and publicize a multi-modal data repository about False News on Weibo Social platform(MCG-FNeWS}) to help evaluate the performance of different approaches from participants.
BibTeX
@misc{cao2019false,
  author = {Juan Cao and Qiang Sheng and Peng Qi and Lei Zhong and Yanyan Wang and Xueyao Zhang},
  title = {{False News Detection on Social Media}},
  year = {2019},
  eprint = {1908.10818},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1908.10818},
}
Figure