Learning from Fact-checkers: Analysis and Generation of Fact-checking Language

Nguyen Vo, Kyumin Lee

Abstract

In fighting against fake news, many fact-checking systems comprised of human-based fact-checking sites (e.g., snopes.com and politifact.com) and automatic detection systems have been developed in recent years. However, online users still keep sharing fake news even when it has been debunked. It means that early fake news detection may be insufficient and we need another complementary approach to mitigate the spread of misinformation. In this paper, we introduce a novel application of text generation for combating fake news. In particular, we (1) leverage online users named fact-checkers, who cite fact-checking sites as credible evidences to fact-check information in public discourse; (2) analyze linguistic characteristics of fact-checking tweets; and (3) propose and build a deep learning framework to generate responses with fact-checking intention to increase the fact-checkers' engagement in fact-checking activities. Our analysis reveals that the fact-checkers tend to refute misinformation and use formal language (e.g. few swear words and Internet slangs). Our framework successfully generates relevant responses, and outperforms competing models by achieving up to 30% improvements. Our qualitative study also confirms that the superiority of our generated responses compared with responses generated from the existing models.

Introduction

While automatic fake news detection has improved significantly, misinformation often persists in digital ecosystems even after being debunked. This phenomenon is largely driven by confirmation bias and the "echo chamber" effect, where users continue to share flagged content within their social circles. Detection alone is a passive defense; the authors argue that active intervention is required to curb the spread of false claims.

One of the most effective countermeasures is direct engagement from "fact-checkers"—ordinary users who reply to misinformation with evidence and links to reputable sites. However, manually crafting these responses is a labor-intensive process that limits the scale of human intervention. This research aims to understand the linguistic patterns of successful fact-checking and introduces the Fact-checking Response Generator (FCRG), a deep learning tool designed to help volunteers scale their impact by generating high-quality, persuasive replies.

Example of a fact-checking intervention on Twitter where a user refutes a false claim with a debunking link.

Linguistic Analysis of Fact-checkers

To build an effective generation model, the researchers first analyzed the linguistic "DNA" of fact-checking tweets (FC-tweets). Using LIWC (Linguistic Inquiry and Word Count) and SEMAFOR semantic parsing, they compared FC-tweets against "Normal Replies" and "Random Replies."

The Fact-Checker’s Persona

Contrary to the stereotype of the "angry internet user," the analysis reveals that fact-checkers maintain a professional and objective persona. Key findings include:

  • Formal Language: Fact-checkers use significantly fewer swear words and less internet slang (netspeak) than average users.
  • Focus on the Claim: They tend to use impersonal pronouns ("this," "it") rather than personal ones, shifting the focus from the individual poster to the validity of the claim itself.
  • Temporal Focus: FC-tweets focus heavily on the past, using past-tense verbs to refer to previously debunked stories or historical evidence, whereas normal replies are more present-oriented.

Semantic Framing

Semantic analysis using SEMAFOR showed that fact-checkers utilize specific frames to categorize misinformation. They frequently use the Artificiality frame (words like "fake," "false," "lie") and the Statement frame (words like "reports," "claims," "said") to distance the claim from reality.

Comparison of semantic frames across different tweet types, highlighting the prevalence of Artificiality and Statement frames in fact-checking.

The FCRG Framework

The authors proposed a neural text generation model based on a Sequence-to-Sequence (Seq2Seq) architecture with a Gated Recurrent Unit (GRU). The framework consists of an encoder to capture the context of the misinformation and a decoder to generate the response.

To improve relevance, the model employs two types of attention mechanisms: Dot Attention (FCRG-DT) and Bilinear Attention (FCRG-BL). These allow the model to focus on specific keywords or claims within the original tweet when generating a reply. Additionally, the model uses shared word embeddings to reduce parameter overhead and prevent overfitting in a domain characterized by specific jargon.

Datasets and Results

The study utilized a curated FC-Tweet Dataset containing 64,110 original tweets and 73,203 associated fact-checking replies, collected via the Hoaxy system.

Quantitative Performance

FCRG models outperformed state-of-the-art baselines like SeqAttB and HRED. In scenarios requiring responses of at least 10 tokens—the "Goldilocks" length found to maximize engagement—FCRG-DT showed a 17.9% improvement in BLEU-3 and a 13% improvement in BLEU-4 over the best baseline.

Human Evaluation and Impact

In blind tests, human evaluators preferred FCRG-DT responses 40% of the time, noting the model's ability to capture specific entities (e.g., "uranium") from the context. Most importantly, the research observed a measurable "chilling effect": after receiving an FC-tweet intervention, 7% of original misinformation tweets were eventually deleted or hidden by the posters.

Conclusion and Future Work

While the FCRG framework marks a significant step toward automated intervention, limitations remain. Currently, the model requires the user to provide the fact-checking URL. Future iterations could integrate automated retrieval systems to fetch debunking articles directly. Furthermore, incorporating the actual evidence text from linked articles—rather than just learning from human commentary—could make generated responses even more persuasive and informative.

Summary updated

Abstract

In fighting against fake news, many fact-checking systems comprised of human-based fact-checking sites (e.g., snopes.com and politifact.com) and automatic detection systems have been developed in recent years. However, online users still keep sharing fake news even when it has been debunked. It means that early fake news detection may be insufficient and we need another complementary approach to mitigate the spread of misinformation. In this paper, we introduce a novel application of text generation for combating fake news. In particular, we (1) leverage online users named \emph{fact-checkers}, who cite fact-checking sites as credible evidences to fact-check information in public discourse; (2) analyze linguistic characteristics of fact-checking tweets; and (3) propose and build a deep learning framework to generate responses with fact-checking intention to increase the fact-checkers' engagement in fact-checking activities. Our analysis reveals that the fact-checkers tend to refute misinformation and use formal language (e.g. few swear words and Internet slangs). Our framework successfully generates relevant responses, and outperforms competing models by achieving up to 30\% improvements. Our qualitative study also confirms that the superiority of our generated responses compared with responses generated from the existing models.
BibTeX
@inproceedings{vo2019learning,
  author = {Nguyen Vo and Kyumin Lee},
  title = {{Learning from Fact-checkers: Analysis and Generation of Fact-checking Language}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2019},
  url = {https://openreview.net/forum?id=kKzq2R8812},
}
Figure