Fact-Checking, Fake News, Propaganda, and Media Bias: Truth Seeking in the Post-Truth Era

Preslav Nakov, Giovanni Da San MartinoEMNLP 2020 · OpenReview

Abstract

In the modern digital landscape, the democratization of content creation has bypassed traditional editorial gatekeepers, leading to a "Post-Truth Era" where emotional appeals often supersede factual accuracy. This paper, based on the EMNLP 2020 tutorial by Preslav Nakov and Giovanni Da San Martino, outlines a comprehensive computational framework for truth-seeking. By framing the fight against disinformation as a multi-stage pipeline—ranging from check-worthiness and stance detection to the analysis of media bias and propaganda—this work provides a unified taxonomy for addressing the global "infodemic."

1. Introduction

The rapid spread of disinformation at "breaking-news speed" has created significant real-world challenges, including election interference and the complexities of the COVID-19 "infodemic." The core problem addressed in this research is the fragmented nature of current solutions. Rather than treating "fake news" as a simple binary classification task, the authors argue for a holistic computational pipeline that identifies, prioritizes, and verifies claims while simultaneously evaluating the reliability of sources and the intent of the actors involved.

2. The Fact-Checking Pipeline

Automated truth-seeking is conceptualized as a sequence of interconnected tasks rather than a monolithic process. The pipeline typically follows this progression:

  • Check-Worthiness Estimation: Utilizing tools like ClaimBuster and ClaimRank, researchers prioritize which statements actually require professional fact-checking, filtering out noise and subjective opinions.
  • Stance Detection: This critical precursor to verification determines whether a specific text supports, denies, or remains neutral toward a given claim.
  • Evidence-Based Verification: Moving beyond simple labels, this stage involves gathering evidence from knowledge bases, Wikipedia, and the broader web to predict veracity based on documented facts.
  • Source Reliability and Media Bias: By shifting focus from individual content pieces to the outlets themselves, models use multi-task learning to estimate the political leaning and trustworthiness of entire news organizations.

3. Propaganda and Malicious Actor Detection

Truth-seeking also requires understanding the rhetorical strategies used to manipulate public opinion. Propaganda detection focuses on identifying specific linguistic "weapons"—such as name-calling, appeals to fear, or loaded language—which may be factually true but are used to manipulate. Furthermore, the identification of social bots and "trolls" is essential to uncovering orchestrated efforts to spread disinformation and distort public discourse.

4. Data Benchmarks

The development of these systems relies on a robust set of benchmarks across various domains:

  • Fact-Checking: FEVER (large-scale evidence extraction), Snopes, ClamsKG, and MultiFC.
  • Fake News and Stance: FakeNewsNet, NELA-GT-2018, and datasets from the Fake News Challenge and SemEval.
  • Media Bias: Data curated from Media Bias Fact/Check, AllSides, and OpenSources.
  • Structured Verification: TabFact, which focuses on verifying claims against structured data tables.

5. Key Insights and the Generative Threat

A significant insight from recent research is the convergence of medical and political misinformation, particularly during the COVID-19 pandemic. This "infodemic" requires models that can bridge domain-specific knowledge with political analysis.

Additionally, the emergence of large-scale language models (such as GPT-2, GPT-3, and GROVER) represents a double-edged sword. While these models can generate high-quality "neural fake news," they also provide the underlying architecture for more sophisticated detection tools. The current state-of-the-art (SOTA) suggests that multi-modal approaches (combining text, images, and social context) and multi-task learning (jointly predicting bias and factuality) are the most promising directions for improvement.

6. Limitations and Future Directions

The battle against disinformation is a perpetual "cat-and-mouse" game. As detection methods improve, so do the techniques for generating sophisticated disinformation (e.g., DeepFakes). Key areas for future work include:

  • Explainability: Moving beyond labels to provide human-readable explanations of why a claim is considered false.
  • Cross-Linguality: Reducing the English-centric nature of current research to address disinformation in non-Western contexts.
  • Real-Time Systems: Bridging the gap between offline dataset evaluation and real-time "journalist-in-the-loop" systems.

7. Conclusion

Combating the post-truth era requires a multi-faceted approach that integrates content verification with source analysis and intent detection. By treating fact-checking as a comprehensive pipeline, the research community can develop more resilient tools to protect the integrity of the global information ecosystem.

Summary updated

BibTeX
@inproceedings{nakov2020fact,
  author = {Preslav Nakov and Giovanni Da San Martino},
  title = {{Fact-Checking, Fake News, Propaganda, and Media Bias: Truth Seeking in the Post-Truth Era}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  url = {https://openreview.net/forum?id=Z1UxcZauuZ},
}
Figure