ARGUS framework studies narrative features' impact on persuasion in argumentative discourse using a new annotated ChangeMyView corpus.
The ARGUS framework investigates how narrative features influence persuasion in online argumentative discourse, specifically within the ChangeMyView Reddit community, by combining a newly annotated corpus with computational models to detect and analyze narrative elements. Grounded in established narrativity theories, ARGUS introduces an annotation scheme that assesses both the presence of a story and six fine-grained narrative features: Agency, Event Sequencing, World Making, Suspense, Curiosity, and Surprise—categorized into text-oriented (structural) and reader-oriented (response) dimensions.
The framework employs encoder-based classifiers (e.g., BERT, RoBERTa) and evaluates their performance against zero-shot large language models (LLMs) like Llama 3.1 in identifying these narrative features. Results show that supervised encoder-based models outperform zero-shot LLMs, and modeling narrativity as a scalar (continuous) property leads to more accurate predictions and downstream insights than binary classifications. Narrativity itself is a strong predictor of persuasive success, with higher narrativity scores significantly increasing the likelihood of receiving a "delta" (Δ), the community's marker for a changed viewpoint.
At the feature level, reader-oriented features—particularly Curiosity and Suspense—have the strongest positive impact on persuasion, with a one-unit increase in the Response dimension raising the odds of persuasion by approximately 80%. In contrast, Surprise has a significant negative effect on persuasion, possibly due to disrupting reader engagement. Structural features like Event Sequencing and Agency show only modest positive contributions. These findings suggest that effective persuasive narratives in online discourse engage readers through information disclosure dynamics rather than relying solely on narrative structure.
The dataset, models, and code are publicly available, supporting further research into narrative persuasion and offering implications for AI applications involving LLMs in generating or analyzing persuasive communication.
ARGUS introduces a novel framework for analyzing the intersection of narrative theory and computational argumentation, specifically investigating how storytelling elements influence the efficacy of persuasive discourse. The study centers on the creation and utilization of a newly annotated version of the ChangeMyView corpus, a popular dataset for studying persuasion. By applying the ARGUS framework, the authors systematically categorize and quantify narrative features—such as the presence of personal anecdotes, emotional arcs, and character development—within argumentative texts. This approach allows for a granular examination of how these stylistic and structural components interact with logical reasoning to persuade an audience.
The research provides key insights by correlating specific narrative features with successful persuasion outcomes, measured by the "Delta" scores in the ChangeMyView dataset. The findings suggest that persuasion is not merely a function of logical soundness or citation strength, but is significantly modulated by the rhetorical packaging of the argument. The authors contribute a valuable resource to the NLP community through their annotated corpus, which establishes a taxonomy for narrative features in debate settings, enabling future research to replicate and expand upon these feature-extraction methods.
This material is particularly relevant to current AI research, especially regarding Large Language Models (LLMs) and their application in persuasive communication. As LLMs become increasingly adept at generating human-like text, understanding the nuances of narrative influence is critical for developing models that can engage in more effective, nuanced, and ethically sound persuasion. The ARGUS framework provides a pathway for integrating narrative intelligence into argumentative systems, moving beyond logic-only approaches to create AI agents that better understand the subtleties of human communication and influence.
The ARGUS framework introduces a systematic approach to studying how narrative features in argumentative texts influence persuasion. Leveraging a newly annotated ChangeMyView corpus, the work investigates structural and stylistic elements—such as storytelling techniques, emotional appeals, and rhetorical devices—that enhance or undermine persuasive effectiveness. By combining computational analysis with linguistic insights, ARGUS provides a structured method to quantify narrative persuasion, bridging gaps between discourse analysis and AI-driven text generation.
Key contributions include: 1. A novel annotation scheme capturing narrative features (e.g., character development, plot structure, and emotional framing) in argumentative discourse. 2. Empirical findings on which narrative techniques correlate with higher persuasion success, informed by statistical modeling of the corpus. 3. Implications for AI systems, particularly LLMs, in generating more effective persuasive content by incorporating narrative strategies.
This research matters because it advances AI-driven persuasion research, offering actionable insights for applications in debate systems, marketing, and misinformation mitigation. By formalizing narrative persuasion, ARGUS enables more controlled experiments in AI-generated argumentation, potentially improving the ethical and strategic use of language models in high-stakes communication.
For further details, see the [arXiv preprint](https://arxiv.org/abs/2602.24109).