StoryScope: Narrative Analysis & Generation

Updated 9 April 2026

StoryScope is a suite of methodologies that combines machine learning, computational narratology, and interactive authoring to analyze and generate narratives.
Its feature-based narrative pipeline employs structured JSON templates and gradient-boosted trees to differentiate human from AI-generated fiction with high accuracy.
StoryScope extends to robust online story detection and real-time adaptive storytelling systems using reinforcement learning and spatial analytics for dynamic narrative delivery.

StoryScope refers to a suite of methodologies, systems, and frameworks for analyzing, detecting, authoring, and dynamically delivering stories by exploiting advances in machine learning, computational narratology, and interactive authoring environments. It encompasses pipelines for feature-rich narrative analysis (notably for distinguishing human from AI-generated fiction), interactive digital storytelling systems, open-ended plot exploration via reinforcement learning, and robust story detection in online discourse. The following sections synthesize the state-of-the-art in StoryScope research, its architectures, modeling paradigms, evaluation metrics, and its principal application domains.

1. Feature-Based Narrative Analysis and Authorship Attribution

StoryScope, as formalized in (Russell et al., 3 Apr 2026), is a three-stage pipeline designed to capture discourse-level narrative features in long-form fiction. This process operates by:

Template Extraction: Transforming stories into structured JSON templates using ten dimensions of the NarraBench taxonomy (Agent, Event, Plot, Structure, Setting, Time, Social Network, Revelation, Perspective, Style). Extraction relies on zero-shot prompting of LLMs (e.g., GPT-5.1).
Comparative Analysis: Comparing sets of six parallel stories—one human-written, five from distinct LLMs—for each writing prompt, extracting per-source divergences and meta-patterns.
Feature Discovery and Encoding: Eliciting 408 candidate features, clustering to 304 (distributed as categorical, ordinal, binary, multi-select, and scale variables), yielding fine-grained narrative representations. The features are encoded via LLMs (e.g., Gemini 3 Flash).

Classification models (e.g., gradient-boosted trees with per-feature Shapley attribution) trained on these features produce strong authorship signals:

Human vs. AI classification achieves 93.2% macro-F₁ (narrative features only) and 68.4% macro-F₁ for 6-way source identification.
A set of 30 "core" features (e.g., thematic explicitness, causal chain continuity, subplots, explicit intertextual reference, temporal complexity) serves as principal separators.

Key findings include:

AI-generated stories display elevated thematic explicitness, moralizing, and streamlined plot structure (single-track, protagonist-driven).
Human-authored stories exhibit greater narrative diversity, temporal complexity, intertextual richness, and moral ambiguity.
Cluster analysis in the induced narrative space demonstrates that human stories occupy rarer, more varied regions, while different LLMs produce tightly overlapping, less distinctive clusters (Russell et al., 3 Apr 2026).

Implications include the potential for durable AI-authorship detection and quantitative proxies for narrative originality grounded in combinatorial feature rarity.

2. Story Detection in Online Communities and Microgenred Texts

StoryScope methodologies have been extended for robust detection of narrative spans within heterogeneous social media corpora. In (Antoniak et al., 2023), the framework operationalizes "story" as a minimal sequence of at least two specific, causally/sequentially related events involving agents, without requiring world-building or genre cues. The process involves:

Construction of balanced, topic-diverse datasets with dense, expert-annotated binary labels, story spans, and event spans (e.g., 502 Reddit posts/comments, 47% positive rate).
Span-based and document-level classification using fine-tuned transformers (RoBERTa), classical models (SVM + TF-IDF), and LLM prompting.
Distillation of salient features: event density, prevalence of past-tense verbs, realis events, first-person singular pronouns, entity/concreteness indices.

Performance on document-level story detection (macro-averaged) reaches up to 93% F₁ (LLMs, span-based RoBERTa approaches), with interpretable feature associations. The system facilitates both broad analytics and fine-grained mapping of "storytelling hotspots" across communities, revealing, for example, that high story rates cluster in support and addiction forums, while policy and religion communities exhibit distinctive distributional signatures.

Applications encompass rhetorical strategy analysis (e.g., measuring the persuasion effect of storytelling in argument forums), content moderation, and cross-platform monitoring of narrativity phenomena (Antoniak et al., 2023).

3. Interactive and Location/Audience-Aware Storytelling Systems

StoryScope principles inform the architecture of dynamic storytelling systems that adapt narrative delivery by integrating real-time sensing, context modeling, and audience segmentation. As detailed in (Burke et al., 2016):

System architecture features per-user wearable devices (e.g., Glass) streaming location, orientation, and POV data, with private downstream channels for text/media delivery.
Middleware supports bidirectional messaging (media, location, branching signals), with control-room applications integrating live video, location telemetry, and human-in-the-loop narrative selection.
Audience adaptation algorithms filter and rank content using spatial relevance $R_{\text{spatial}}(u,c) = \exp(-d(L(u), \text{loc}(c))/\sigma)$ , persona-role compatibility $R_{\text{role}}(u,c)$ , and sentiment-driven content targeting (emotionally classified tweet selection with multi-factor objectives).
Story Progression Loop combines real-time author/ML-driven selection of script elements, dynamic delivery, user-audience reaction monitoring, and adaptive branching.

Case studies ("Grace Plains," "Bodies for a Global Brain") demonstrate high engagement, complex branching, and system responsiveness (<500 ms round trip), confirming the utility of hybrid (human + ML) real-time authorship, per-user channeling, and instrumented low-latency telemetry. Lessons for scalable StoryScope deployments include modularity, author-tool autonomy as branching increases, reinforcement-learning-based branch pruning, and adaptive user preference modeling.

4. Visual Storytelling, Authoring Environments, and Scrollytelling

StoryScope has been instantiated in web-based visual authoring systems for scroll-driven ("scrollytelling") interactive narratives (Mörth et al., 2022). Its architecture comprises:

Authoring Environment: No-code, drag-and-drop tools ("Node-Link Canvas") for importing multimodal assets (text, images, video, mesh, maps) and configuring content nodes, parameterization, sequencing, branching, and anchor mapping.
Runtime Engine: Static-site deployment; browser-based engine interprets the story graph, segment/anchor positions, and scroll position $s$ , mapping to normalized segment parameters $t(s)$ and visual interpolation (e.g., $p(s) = p_0 + (p_1 - p_0) f(t(s))$ , $f$ smoothstep/easing).
Visualization Integration: Out-of-the-box support for Mapbox maps, Three.js 3D models, volume rendering, with scroll-driven UI hooks and declarative parameter transition.
Export Model: Generates standalone HTML/CSS/JS builds with bundled assets, suitable for static web hosting (e.g., GitHub Pages, S3).

Evaluation with expert users and domain practitioners demonstrates rapid story composition (mean ≈5 min 14 s/story), high SUS usability (mean 90.6), and demand for features such as math/LaTeX nodes, collaborative editing, and mobile responsiveness. The system's modularity, support for branching, and capability for interactive visualization are highlighted as key affordances for news, education, and scientific communication (Mörth et al., 2022).

5. Reinforcement Learning and Search-Based Narrative Generation

Reinforcement learning frameworks, exemplified by COS(M+O)S (Materzok, 28 Jan 2025), apply StoryScope concepts to systematic exploration of story space through Monte Carlo Tree Search (MCTS) and RL fine-tuning:

Optimization Algorithm: Each story node $s$ is a partial plot; actions $a$ expand the story via LLM-governed chain-of-thought instructions. At every node, multiple actions are generated, each leading to a new story continuation.
MCTS Procedure: Uses UCB-based selection $a^* = \arg\max_a [Q(s,a) + c\sqrt{\ln N(s)/N(s,a)}]$ , policy-guided action sampling, and deferred evaluation (post-50% completion) via a step-level value function $V(s)$ .
Value Model: $R_{\text{role}}(u,c)$ 0 incorporates a curiosity reward (inverted-U of token surprisal) and coherence (average pairwise sentence embedding cosine similarity), emphasizing moderate unpredictability and narrative consistency.
Policy Improvement: Preference optimization via Odds-Ratio Preference Optimization (ORPO), combining supervised loss on preferred actions and log-odds upweighting. Policy is fine-tuned (QLoRA, rank-64 adapters) per high-value trajectories.
Results: Small (3B parameter) LLMs, when combined with search and RL, approach the plot quality of much larger (70B) LLMs on short-story tasks, with >67% human preference for MCTS-optimized expansions over uninformed policies and no statistically significant difference from 70B models on paired GPT-4o evaluations (p=0.93).

This suggests that structured search and reward shaping can compensate for LLM size limitations in creative narrative tasks, at least within moderate-length story domains. Extensions under discussion include value model enrichment, hierarchical search, parallelization, and more granular reference tracking (Materzok, 28 Jan 2025).

6. Applications, Limitations, and Future Directions

Collectively, StoryScope paradigms support a spectrum of research and practical applications:

Detection and attribution of AI-authored fiction, with emphasis on robust narrative signals beyond surface style (Russell et al., 3 Apr 2026).
Fine-grained mapping and analysis of storytelling prevalence and function in online communities, enabling both social research (e.g., persuasion, trust patterns) and platform moderation (Antoniak et al., 2023).
Real-time, audience-adapted, and location-aware narrative delivery, facilitating interactive theater, immersive education, or data journalism (Burke et al., 2016, Mörth et al., 2022).
Systematic plot discovery and generation in constrained or low-resource creative environments, leveraging RL and tree-search techniques (Materzok, 28 Jan 2025).

Known limitations include annotation bottlenecks for narrative features, manual curation burdens in highly branching interactive systems, scaling challenges in RL-based search for long narratives, and open challenges in generalizing feature sets or story detection models across platforms and modalities. Advancements in modular toolsets, semi-supervised annotation, reinforcement learning curricula, and universal narrativity schemas are likely to drive further progress.

References: