Automated rhetorical move and step recognition in fact-checking articles with neural models

As online misinformation has drawn social concerns, extensive efforts have been dedicated to fact-checking, which can help contain the spread of misinformation. Among them, persuasive fact-checking articles play a fundamental role, but little work has focused on their discourse structures that are important for understanding how they work. Rhetorical moves and steps, common in genre analysis, can be used to figure out text structures and communicative goals. Based on existing literature, this research first summarizes a rhetorical structure comprising five moves and six steps for fact-checking articles, which describes how they are organized to achieve the persuasive purpose. We then produce a corpus including 420 articles with annotations of our structures. For automated recognition, we propose our BiLSTM with Hierarchical Attention model, which achieves micro-F1 scores of 70 % and 61.5 % for moves and steps, respectively. The performance and subsequent ablation study demonstrate the effectiveness of our model. Utilizing it, an analysis of the distribution and patterns of moves and steps is conducted on an expanded set of 3800 fact-checking articles. Accordingly, we find that the distributions of rhetorical structures in articles have common characteristics and unique differences, reflecting the strategies used when writing. We further conducted sequence mining, and the obtained frequent sequences can help improve fact-checking writing and provide new ideas for studying the relationship between fact-checking texts and their persuasive effects. Generally, the fact-checking rhetorical structures and the automated model proposed in this work have the potential to help leverage the fact-checking corpus and finally contribute to rebutting misinformation.