How Stories Spread: Modeling Online Narrative Response

Author: Denis Avetisyan

New research introduces a framework for understanding how readers interpret and share stories across different social media platforms.

SocialStoryFrames formalize the reasoning behind both the intent embedded within conversational stories on social media and the reception of those stories by an audience.

This paper presents Social Story Frames, a formalism and pipeline for comparative analysis of narrative practices and reader response using computational social science methods.

While stories powerfully shape our understanding and responses, computational models struggle to capture the nuanced inferences readers make about narrative intent and emotional impact. This paper introduces Social Story Frames: Contextual Reasoning about Narrative Intent and Reception, a formalism and modeling pipeline designed to distill plausible reader responses – including perceived author intent, affective reactions, and value judgements – from conversational context. By linking fine-grained analysis with a taxonomy grounded in narrative theory, we enable comparative study of storytelling practices across online communities. How might a deeper understanding of these shared narrative frames illuminate the dynamics of social interaction and belief formation online?

The Challenge of Narrative Reception: A Formal Inquiry

The efficacy of communication fundamentally rests on anticipating how an audience will interpret a given message, and this is particularly true for narratives. However, predicting narrative reception presents a significant challenge, extending beyond simple message transmission. Stories don’t carry meaning passively; instead, meaning is actively constructed by the recipient, shaped by a complex interplay of factors. These include pre-existing beliefs, emotional state, cultural background, and even momentary attention levels. Consequently, even meticulously crafted narratives can be misinterpreted, leading to unintended consequences or a failure to connect with the intended audience. Researchers continue to grapple with the variables influencing this process, striving to move beyond generalized assumptions and towards more precise models of how audiences engage with and ultimately understand stories.

Historically, analyzing how stories are received has been hampered by a tendency to prioritize either the author’s intended message, the surrounding cultural climate, or the reader’s personal biases in isolation. This simplification neglects the intricate dance between these elements; a narrative’s impact isn’t solely dictated by what the creator wants to convey, nor is it simply a product of the reader’s pre-existing beliefs. Instead, meaning emerges from a complex interaction where contextual factors – encompassing everything from historical events to immediate social settings – actively shape how a reader interprets the narrative’s intent. Consequently, understanding reception requires moving beyond singular viewpoints and embracing a more holistic model that acknowledges the dynamic and reciprocal relationship between story, context, and individual perception.

Analysis of subreddits reveals diverse narrative practices viewed through author-focused and reader-focused perspectives.

SocialStoryFrames: A Formalization of Narrative Analysis

SocialStoryFrames (SSF) addresses the inherent complexity in narrative communication by providing a structured methodology for analyzing reader reception. Traditional approaches often treat interpretation as subjective and unpredictable; SSF, however, posits that responses are patterned and can be systematically studied. This is achieved through the formalization of analytical processes, moving beyond qualitative assessments to quantifiable data regarding how audiences engage with narratives. By explicitly modeling the transition from an author’s intended message to a reader’s constructed understanding, SSF facilitates a more rigorous examination of the factors influencing interpretation and allows for predictive modeling of audience response.

The SocialStoryFrames (SSF) framework utilizes a structured, hierarchical taxonomy – the SSF-Taxonomy – to systematically categorize and analyze the spectrum of reader responses to narratives. This taxonomy is not a flat list of reactions, but rather a multi-level classification system allowing for granular distinctions between response types. Higher levels represent broad categories of interpretation, while successive levels detail specific facets of those interpretations – encompassing cognitive, emotional, and behavioral reactions. The SSF-Taxonomy enables researchers to move beyond simple positive/negative sentiment analysis and identify why audiences respond in particular ways, facilitating a more nuanced understanding of narrative reception and allowing for comparative analysis across diverse reader groups and story types.

SocialStoryFrames (SSF) prioritizes contextual analysis as a core component of understanding narrative reception because interpretation is demonstrably influenced by external factors. These factors include, but are not limited to, the reader’s pre-existing beliefs, the social environment in which the story is encountered, and concurrent events impacting the audience. By systematically incorporating these contextual variables into its analytical process, SSF aims to move beyond purely textual analysis and generate predictions about likely audience responses with increased accuracy. This contextual awareness allows for the differentiation between intrinsic narrative elements and extrinsic influences on interpretation, improving the reliability of analytical outcomes and providing a more nuanced understanding of how stories function within real-world settings.

We present SSF-Taxonomy, a novel categorization of how readers respond to informal stories shared on social media.

Computational Models for Narrative Inference: SSF-Generator and SSF-Classifier

SSF-Generator is a language model built upon the Llama-3.1 architecture, specifically designed to generate plausible continuations or responses to provided narrative prompts. This model utilizes the capabilities of Llama-3.1 to predict likely subsequent events or character actions within a given story context. The generation process relies on the model’s learned statistical relationships within the training data to produce coherent and contextually relevant inferences. SSF-Generator’s output is not limited to single-answer predictions; it is capable of producing a range of possible responses, facilitating analysis of diverse storytelling approaches and community variations. The model’s performance is evaluated based on metrics assessing both the plausibility and coherence of the generated text, ensuring the inferences align with established narrative conventions and the provided input.

SSF-Classifier is a supervised finetuned model designed to categorize the responses generated by SSF-Generator according to the predefined SSF-Taxonomy. This taxonomy provides a structured framework for analyzing narrative inferences, enabling consistent and quantifiable assessment. Supervised finetuning involved training the classifier on a labeled dataset of narrative responses mapped to specific categories within the SSF-Taxonomy, optimizing its ability to accurately assign categories to newly generated inferences. The resulting model facilitates automated analysis of storytelling variations by providing categorical labels for each response, allowing for quantitative comparisons across different narratives and communities.

To quantify the relationship between generated narrative inferences, both Cosine Similarity and Jensen-Shannon Divergence (JSD) were employed. Cosine Similarity measures the angle between vector representations of the inferences, providing a value between -1 and 1, where higher values indicate greater similarity. JSD, calculated from the probability distributions of token usage within the inferences, provides a distance metric between 0 and 1, with lower values indicating greater similarity. Our SSF-Sim tool utilizes these metrics to assess both the consistency of individual inferences and the overall diversity of responses across different narrative prompts. Specifically, analysis of JSD values reveals statistically significant variations in storytelling practices between the communities represented in our dataset, demonstrating the tool’s capability to provide measurable insights into cultural differences in narrative construction. $JSD(P||Q) = \frac{1}{2}D_{KL}(P|Q) + \frac{1}{2}D_{KL}(Q|P)$ , where $D_{KL}$ is the Kullback-Leibler divergence.

GPT-4o and the distilled SSF-Generator effectively produce plausible inferences regarding reader responses, as indicated by high ratings of likelihood.

The SSF-Corpus: Grounding Analysis in Empirical Data

The creation of a comprehensive and carefully curated dataset, the SSF-Corpus, was foundational to the development of robust and generalizable narrative understanding models. This corpus, comprising a substantial collection of stories paired with diverse reader interpretations, serves as the primary training ground and evaluation benchmark. By exposing models to a wide spectrum of responses – varying in length, detail, and perspective – the SSF-Corpus facilitates the learning of nuanced relationships between narrative content and subjective interpretation. The scale and diversity of the corpus are critical; it moves beyond limited datasets and enables the models to perform reliably across a broader range of storytelling and reader profiles, ultimately enhancing their ability to accurately predict and analyze how different individuals engage with the same narrative.

The SSF-Corpus, through its collection of diverse reader responses to shared narratives, illuminates the remarkable spectrum of interpretation possible even with identical source material. Analysis demonstrates that individuals frequently emphasize different elements of a story, leading to varied emotional reactions, thematic understandings, and recall details. This isn’t simply random noise; the corpus reveals patterns in these divergences, suggesting that factors like personal experience, cultural background, and even reading style systematically influence how a story is processed and internalized. Consequently, the corpus doesn’t just offer data; it provides a window into the cognitive processes underlying narrative comprehension and the inherently subjective nature of meaning-making, highlighting that a single story can truly hold multiple truths for different readers.

The nuanced nature of story interpretation necessitates a robust method for gauging the relatedness of different reader responses. Recent findings highlight that standard semantic similarity metrics often fall short in capturing the subtle connections between interpretations, even when those interpretations share core meaning. To address this, researchers developed SSF-Sim, a novel approach that demonstrably outperforms these conventional metrics – as visualized in Figure 11 – by focusing on semantic relationships specific to narrative understanding. This improved ability to discern meaningful connections between interpretations allows for a more accurate assessment of reader response diversity and a deeper understanding of how individuals construct meaning from the same story, paving the way for more sophisticated analyses of narrative engagement.

SSF-Sim more accurately reflects human preferences for story similarity compared to standard semantic similarity metrics.

Towards a Formalized Science of Narrative: Future Directions

SocialStoryFrames establishes a robust framework for investigating how stories are constructed and shared within diverse social groups. This system moves beyond universal narrative structures to recognize the nuanced ways communities shape their accounts of events, experiences, and beliefs. By identifying and categorizing recurring elements – such as plot drivers, character roles, and moral lessons – across a broad spectrum of narratives, researchers gain the tools to compare storytelling practices between cultures and pinpoint variations linked to specific social contexts. The framework isn’t simply a catalog of narrative components; it’s a methodology for understanding how these components are assembled and what those assemblies reveal about the values, norms, and worldviews of the communities generating them, fostering a deeper appreciation for the cultural specificity of human storytelling.

The continued development of the SocialStoryFrames (SSF) Corpus represents a crucial next step in understanding the nuances of storytelling across varied human experiences. Researchers plan to significantly broaden the corpus’s scope, moving beyond initial datasets to incorporate narratives from a wider array of cultural backgrounds, geographic locations, and individual perspectives. This expansion isn’t simply about quantity; the aim is to capture the rich tapestry of human communication, including traditionally underrepresented voices and storytelling styles. By increasing the diversity of narratives within the SSF-Corpus, the framework promises to yield more robust and generalizable insights into the fundamental building blocks of stories and how meaning is constructed and shared, ultimately fostering a more inclusive and accurate model of narrative practices globally.

The development of a robust framework for analyzing narrative practices opens avenues for diverse applications, extending beyond theoretical linguistics into practical fields. Researchers anticipate leveraging these insights to craft more engaging and personally relevant storytelling experiences, potentially revolutionizing entertainment and educational content. Moreover, a deeper understanding of how narratives are constructed and interpreted promises to refine communication strategies across various sectors, including marketing, healthcare, and diplomacy. The consistent achievement of a Macro F1 Score of 0.80 or higher throughout the annotation refinement process underscores the dependability and validity of the analytical guidelines, bolstering confidence in the framework’s ability to consistently and accurately categorize narrative elements across different contexts and communities.

Annotators were presented with a story and associated conversational context to begin each annotation task.

The pursuit of understanding how narratives resonate – or fail to – with different audiences, as detailed in this work on Social Story Frames, echoes a fundamental principle of logical construction. One might observe that, as Alan Turing stated, “Sometimes people who are unskillful learners do not ask very good questions.” The SSF formalism attempts to elicit the ‘good questions’ – the underlying assumptions and contextual reasoning – that govern reader reception. By explicitly modeling these frames, the system moves beyond simply observing response to understanding the invariant principles driving it. The SSF pipeline doesn’t merely document what readers think, but strives to reveal why they think it, mirroring a desire for provable understanding rather than empirical observation alone.

What’s Next?

The introduction of Social Story Frames represents a necessary, if not entirely sufficient, attempt to formalize the inherently messy process of narrative reception. The formalism itself is elegant; the pipeline, functional. Yet, the true challenge lies not in modeling response, but in achieving a provable consistency between computational prediction and the subjective experience of understanding. Current evaluations, reliant on aggregate social media metrics, offer only statistical correlations – a far cry from demonstrating genuine comprehension.

Future work must address the limitations of relying solely on observed behavior as a proxy for internal cognitive states. The field requires a move toward grounding these frames in established theories of narrative psychology and cognitive science. Specifically, the model’s ability to account for individual differences in interpretation-the subtle variations in schema, belief, and emotional state-remains underdeveloped. A rigorous examination of edge cases-stories deliberately crafted to mislead, or those relying on deeply contextual knowledge-will be critical.

Ultimately, the utility of Social Story Frames, or any similar formalism, will be determined not by its ability to predict what is shared, but by its capacity to reveal why. Until the model can articulate the logical steps by which a narrative elicits a specific response-a response verifiable against a principled understanding of cognition-it remains a sophisticated, but incomplete, approximation of a fundamentally human process.

Original article: https://arxiv.org/pdf/2512.15925.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/