Where Are the Facts? Searching for Fact-checked Information to Alleviate the Spread of Fake News

Nguyen Vo, Kyumin Lee

Abstract

Although many fact-checking systems have been developed in academia and industry, fake news is still proliferating on social media. These systems mostly focus on fact-checking but usually neglect online users who are the main drivers of the spread of misinformation. How can we use fact-checked information to improve users' consciousness of fake news to which they are exposed? How can we stop users from spreading fake news? To tackle these questions, we propose a novel framework to search for fact-checking articles, which address the content of an original tweet (that may contain misinformation) posted by online users. The search can directly warn fake news posters and online users (e.g. the posters' followers) about misinformation, discourage them from spreading fake news, and scale up verified content on social media. Our framework uses both text and images to search for fact-checking articles, and achieves promising results on real-world datasets.

Introduction

The proliferation of misinformation on social media remains a critical challenge despite advancements in automated fact-checking. A significant gap exists between detecting fake news and actively informing the users who consume and share it. Often, users spread misinformation not out of malice but due to a lack of awareness that a claim has been debunked. To address this, we propose a system that automatically retrieves existing fact-checking (FC) articles to directly verify or debunk specific social media posts.

The Multimodal Misinformation Challenge

Misinformation on platforms like Twitter is rarely confined to plain text. It frequently utilizes "fauxtography"—misleading images paired with fake quotes or captions. In these instances, the tweet's text may be generic or context-dependent, while the actual claim is embedded within the image itself. Standard text-based search engines fail to capture this context.

Figure 1: A social media post where the core claim is embedded within an image rather than the text.

As shown in Figure 1, the tweet text "Keep your promise Barack" provides little information about the specific claim. The actual misinformation—a fake CNN headline—is contained within the image. Our framework addresses this by using Optical Character Recognition (OCR) to extract text from images, expanding the search query to include the full multimodal context.

Methodology: The Multimodal Attention Network (MAN)

The proposed framework operates in two stages: candidate retrieval and neural re-ranking. Initial retrieval uses the BM25 algorithm on a query expanded by OCR-extracted text. The core innovation lies in the Multimodal Attention Network (MAN), which re-ranks these candidates.

Figure 2: The architecture of the Multimodal Attention Network (MAN), showing textual and visual matching layers.

The MAN architecture (Figure 2) consists of four primary layers:

  1. Projection Layers: Utilizes both static (Glove) and contextual (ELMo) word embeddings to represent text.
  2. Textual Matching Layer: Constructs interaction matrices between the tweet and the article. It employs a novel attention mechanism where match weights are determined by the Euclidean distance between contextual ELMo embeddings, allowing for "soft-matching" of typos (e.g., "Hillar") and slang.
  3. Visual Matching Layer: Employs a pre-trained ResNet50 to compare visual similarities between images in the tweet and those in the fact-checking article.
  4. Unifying Layer: Fuses textual and visual signals to produce a final relevancy score.

Experimental Results

The framework was evaluated on two datasets centered on the 2016 U.S. Presidential election: Snopes (11,202 pairs) and PolitiFact (2,037 pairs).

Key Findings

  • Impact of Visual Context: Incorporating OCR-extracted text significantly boosted retrieval performance. On the Snopes dataset, HIT@50 improved from 50% to nearly 90% when image text was included.
  • Model Performance: The MAN model consistently outperformed nine state-of-the-art baselines across all metrics (NDCG and HIT@K). In text-sparse scenarios, MAN showed an average improvement of 17.2% over the best baselines.
  • Real-World Impact: By accurately retrieving FC articles for a small test set of 910 tweets, the system could have potentially reached 233 million followers, preventing an estimated half a million shares.

Conclusion

This research demonstrates that a multimodal approach is essential for effectively retrieving fact-checking information on social media. While the current system relies on text for initial retrieval, future work will explore incorporating image-to-image similarity at the retrieval stage to handle purely visual misinformation. By closing the loop between fact-checkers and social media users, we can more effectively mitigate the spread of fake news.

Summary updated

Abstract

Although many fact-checking systems have been developed in academia and industry, fake news is still proliferating on social media. These systems mostly focus on fact-checking but usually neglect online users who are the main drivers of the spread of misinformation. How can we use fact-checked information to improve users' consciousness of fake news to which they are exposed? How can we stop users from spreading fake news? To tackle these questions, we propose a novel framework to search for fact-checking articles, which address the content of an original tweet (that may contain misinformation) posted by online users. The search can directly warn fake news posters and online users (e.g. the posters' followers) about misinformation, discourage them from spreading fake news, and scale up verified content on social media. Our framework uses both text and images to search for fact-checking articles, and achieves promising results on real-world datasets. Our code and datasets are released at https://github.com/nguyenvo09/EMNLP2020.
BibTeX
@inproceedings{vo2020where,
  author = {Nguyen Vo and Kyumin Lee},
  title = {{Where Are the Facts? Searching for Fact-checked Information to Alleviate the Spread of Fake News}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  url = {https://openreview.net/forum?id=QuLyQtzsOC},
}
Figure