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 fake news on social media remains a critical challenge despite advancements in automated fact-checking. Most existing research focuses on detection—classifying a post as true or false. However, detection alone does not necessarily curb the spread of misinformation. To effectively change user behavior, it is necessary to provide credible, evidence-based rebuttals.

A significant gap exists: many claims shared on social media have already been debunked by professional fact-checking organizations like Snopes or PolitiFact. However, users continue to share misinformation because they are unaware of these existing fact-checks. The core challenge is one of retrieval: mapping messy, multimodal social media posts (often containing memes with minimal captions) to long-form, authoritative fact-checking articles.

Methodology: The Multimodal Attention Network (MAN)

To address this retrieval challenge, we propose a two-stage framework. The first stage involves Basic Retrieval (BM25-TI), which utilizes the BM25 algorithm enhanced by Optical Character Recognition (OCR). Since many "fauxtography" posts embed claims within images, OCR extracts this text to enrich the search query.

The second stage is the Multimodal Attention Network (MAN), a re-ranking model that performs deep cross-modal comparisons between the tweet and candidate articles.

The Multimodal Attention Network (MAN) architecture for matching tweets to fact-checking articles.

As shown in Figure 3, the MAN model operates through two primary channels:

  1. Textual Matching: The model integrates static (GloVe) and contextual (ELMo) embeddings. It uses the distance between contextual embeddings to weight keyword matches, ensuring that words are compared within their specific semantic contexts (e.g., distinguishing between a financial "bank" and a river "bank").
  2. Visual Matching: Using a pre-trained ResNet50, the model extracts features from images in both the tweet and the fact-check article, calculating pairwise similarities to identify visual relationships.

These signals are fused through a neural network to generate a final relevancy score, allowing the system to rank the most pertinent fact-checking articles for a given tweet.

Experimental Results

We evaluated our framework on a newly curated dataset of over 13,000 tweet-article pairs sourced from Snopes and PolitiFact. The results demonstrate that the MAN model significantly outperforms state-of-the-art baselines in retrieval accuracy.

Comparison of retrieval performance (HIT@K) on Snopes and PolitiFact datasets.

Key findings include:

  • Performance Gains: On the Snopes dataset, the MAN model achieved a 39.6% improvement in NDCG@1 over the best baseline.
  • Impact of OCR: Figure 6 illustrates the effectiveness of combining text and image-extracted text (BM25-TI). On the Snopes dataset, adding OCR-extracted text increased the Hit Rate (finding the correct article in the top 50) from 50% to nearly 90%.
  • Real-World Scalability: Analysis suggests that applying this model to the test set could have informed approximately 233 million followers, potentially preventing over half a million retweets of misinformation.

Conclusion and Future Work

By shifting the focus from detection to retrieval, we can leverage the existing body of professional fact-checks to proactively combat fake news. While the MAN model shows strong performance, limitations remain. The initial retrieval stage still relies on text; future work could explore "Visual BM25" for purely image-based queries. Additionally, extending the framework to handle video content and "Deepfakes" is a critical next step in addressing the evolving landscape of digital misinformation.

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