ARGUS: Seeing the Influence of Narrative Features on Persuasion in Argumentative Texts
| Authors | Sara Nabhani et al. |
| Year | 2026 |
| Field | NLP |
| arXiv | 2602.24109 |
| Download | |
| Categories | cs.CL, cs.AI |
Abstract
Can narratives make arguments more persuasive? And to this end, which narrative features matter most? Although stories are often seen as powerful tools for persuasion, their specific role in online, unstructured argumentation remains underexplored. To address this gap, we present ARGUS, a framework for studying the impact of narration on persuasion in argumentative discourse. ARGUS introduces a new ChangeMyView corpus annotated for story presence and six key narrative features, integrating insights from two established theoretical frameworks that capture both textual narrative features and their effects on recipients. Leveraging both encoder-based classifiers and zero-shot large language models (LLMs), ARGUS identifies stories and narrative features and applies them at scale to examine how different narrative dimensions influence persuasion success in online argumentation.
Engineering Breakdown
Plain English
This paper introduces ARGUS, a framework for understanding how narrative elements (stories) affect persuasiveness in argumentative online text, specifically using Reddit's ChangeMyView dataset. The authors created a new corpus annotated with story presence and six key narrative features, then built classifiers (both encoder-based and zero-shot LLMs) to identify these elements at scale and measure their impact on persuasion outcomes. The core finding is that narratives aren't just present in arguments—specific narrative features have measurable effects on whether people actually change their minds. This bridges the gap between storytelling research and computational argumentation, providing concrete evidence about which narrative techniques matter most in real online debates.
Core Technical Contribution
ARGUS's main innovation is a theoretically-grounded annotation scheme that combines textual narrative feature detection with persuasion outcome measurement in a single framework. Rather than treating narrative as a binary presence/absence question, the authors identified six specific narrative features (derived from established narrative theory) and measured their independent contributions to persuasion. The technical novelty lies in applying both supervised encoder-based classifiers and zero-shot LLMs to identify these fine-grained features in unstructured argumentative text, then using this labeled data to study causal relationships between narrative presence and argument success. This represents the first systematic, large-scale study linking narrative features directly to persuasion outcomes in real argumentative discourse.
How It Works
The ARGUS pipeline operates in three stages: annotation, feature detection, and persuasion analysis. First, human annotators label ChangeMyView posts and comments for story presence and six narrative features (the paper doesn't specify all six, but examples likely include narrative coherence, personal experience framing, temporal sequencing, and emotional elements). Second, the framework trains encoder-based classifiers (transformer models like BERT or RoBERTa) on this labeled data to automatically identify stories and each narrative feature in new text, while also experimenting with zero-shot capabilities from LLMs to reduce annotation burden. Third, the system correlates the presence and strength of each narrative feature with persuasion outcomes (measured by whether the original poster declared a changed view), enabling statistical analysis of which features correlate most strongly with successful persuasion. The output is both a feature-annotated corpus and empirical metrics showing feature-to-persuasion impact strength.
Production Impact
For teams building persuasion-aware systems—recommendation engines, content moderation, debate platforms, or argument generation tools—ARGUS provides a reusable feature extraction pipeline that identifies which narrative elements drive engagement and outcome changes. Production engineers could integrate this framework to: (1) automatically score incoming argumentative content for narrative persuasiveness, enabling smarter ranking or recommendation; (2) provide feedback to content creators about which narrative techniques strengthen their arguments; (3) detect manipulative narrative framing at scale. The trade-off is that you need labeled training data (the corpus), and inference latency depends on whether you use lightweight encoder classifiers (faster, ~50-100ms per text) or zero-shot LLMs (slower, higher cost). The framework is particularly valuable for platforms where persuasion is a direct outcome metric, but less useful in contexts where narrative isn't a primary factor.
Limitations and When Not to Use This
ARGUS is constrained to the ChangeMyView domain, where arguments explicitly conclude with binary persuasion outcomes (poster changed mind or didn't)—this may not generalize to other argumentative contexts like legal briefs, scientific papers, or casual debates where persuasion signals are implicit or absent. The paper doesn't specify the six narrative features in the abstract, making it unclear how universally applicable they are or whether they're specific to online discussion rhetoric. Zero-shot LLM performance on narrative feature detection is mentioned but not benchmarked, so there's uncertainty about whether those models truly capture the features without task-specific training. Finally, the framework measures correlation between narrative features and persuasion, not causation—it's possible that better arguments simply include more narratives, rather than narratives themselves driving persuasion.
Research Context
This work bridges narrative analysis (established in literary criticism and psychology) with computational argumentation and NLP persuasion detection. It builds on prior work in argument mining and persuasion detection (which typically focuses on structural or linguistic features) but adds a narrative dimension that has been underexplored in online argumentation research. The ChangeMyView dataset has become a standard benchmark for persuasion research, and this paper extends it with richer annotations (narrative layers) that enable more nuanced analysis. The research opens directions for studying whether narrative techniques can be taught or induced in argument generation systems, and whether different narrative features work better for different argument types or audience segments.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::