Automated Fact-Checking: A New Study on Information Technology

A team of researchers from Nanjing University has published a groundbreaking study on fact-checking in the field of information technology. The study aimed to develop a method for automatically recognizing persuasive fact-checking articles and providing insights into their discourse structures. Using a corpus of 420 annotated articles and a BiLSTM with Hierarchical Attention model, the researchers achieved micro-F1 scores of 70% and 61.5% for moves and steps, respectively. Their findings suggest that fact-checking rhetorical structures and automated models can help leverage fact-checking corpora and contribute to rebutting misinformation.

Key Takeaways:

The study highlights the importance of persuasive fact-checking articles in containing the spread of misinformation online, emphasizing the need for a deeper understanding of their discourse structures.
Researchers developed a rhetorical structure comprising five moves and six steps for fact-checking articles, which describes how they achieve persuasive purposes.
A corpus of 420 articles was annotated with the proposed structure, and a BiLSTM with Hierarchical Attention model was proposed for automated recognition, achieving significant results.
The study demonstrated the effectiveness of the proposed model through ablation study and further analysis of the distribution and patterns of moves and steps in a larger set of 3800 fact-checking articles.
Frequent sequences obtained through sequence mining can help improve fact-checking writing and provide new ideas for studying the relationship between fact-checking texts and their persuasive effects.
The research concluded that fact-checking rhetorical structures and automated models have the potential to help leverage fact-checking corpora and contribute to rebutting misinformation.

Statistics:

420 articles were annotated with the proposed rhetorical structure for fact-checking articles.
The BiLSTM with Hierarchical Attention model achieved micro-F1 scores of 70% and 61.5% for moves and steps, respectively.
After ablation study, the model's performance was tested on an expanded set of 3800 fact-checking articles.
Sequence mining obtained frequent sequences that can help improve fact-checking writing and provide new ideas for studying the relationship between fact-checking texts and their persuasive effects.
62% of fact-checking rhetorical structures in articles have common characteristics, while 38% exhibit unique differences.

Sources:

Automated Rhetorical Move and Step Recognition In Fact-checking Articles With Neural Models. Information Processing & Management, 2025; 62(6).
Information Processing & Management. Elsevier Sci Ltd, 125 London Wall, London, England.
Ningyuan Song, Nanjing University, School of Information Management, 163 Xianlin Rd, Nanjing 210023, People's Republic of China.