StoryExplorer: A Visualization Framework for Storyline Generation of Textual Narratives

Li Ye, Lei Wang, Shaolun Ruan, Heyu Wang, Yuwei Meng, Yigang Wang, Wei Chen, Zhiguang Zhou

TL;DR

This paper introduces StoryExplorer, a novel visualization framework designed to help users extract and understand storylines from large volumes of narrative text. By combining interactive annotation, GPT-powered entity recognition, and a structured workflow, StoryExplorer aims to reduce cognitive load and improve the formation of coherent mental models of narrative content.

What Problem Does This Paper Address?

In today's digital age, we're inundated with narrative texts like novels and news articles. While these texts contain rich information, their sheer volume and complexity make it incredibly difficult for readers to extract and retain the core storyline. This challenge stems from limitations in human working memory and attention span, leading to information overload and reduced comprehension efficiency. Existing text-mining techniques often fall short in producing intuitive and personalized story structures, necessitating a more effective approach for knowledge externalization.

Methods

StoryExplorer tackles this problem with a three-stage workflow: "insight finding-scripting-storytelling."

  • Insight Finding (Selection): Users interact with the text using a "Text View." They can highlight key sentences and use stroke annotations to mark entities. Crucially, StoryExplorer integrates GPT-based visual hints to automatically recognize and suggest entities (people, places, events, time), significantly reducing the manual effort of identification.
  • Scripting (Organization): In the "Fragment View," users organize these identified entities into story fragments. This stage allows for classifying entities, adding keywords, and generating brief text summaries for each fragment. A word cloud feature helps users quickly grasp the essence of a fragment.
  • Storytelling (Integration): The "Storyline View" is where users assemble these fragments into a coherent storyline. StoryExplorer supports interactive editing, allowing users to adjust fragment placement, merge or delete fragments, and visualize temporal relationships. The system also offers features like zooming, panning, and a mini-map for navigation, facilitating both overview and detailed exploration.

The system is implemented as a web-based client-server application using Flask, with the client side handling annotations and rendering SVG elements.

Datasets

The paper does not explicitly list or introduce new datasets for training. Instead, it uses existing narrative texts for its case studies and user evaluations. Two primary texts are mentioned:

  • "Harry Potter and the Sorcerer's Stone" (first three chapters)
  • "Pride & Prejudice"
  • "The Tinder-Box" (used in the formative study)
  • "Schindler's List" (script text)

Key Insights and Ideas

  • Bridging Working Memory and Long-Term Memory: The core idea is to externalize narrative structures in a visual and interactive way to overcome the limitations of human working memory, thereby facilitating the formation of more robust long-term mental models.
  • Human-in-the-Loop with AI Assistance: StoryExplorer doesn't aim for full automation. It intelligently combines user-driven annotation (stroke, highlight) with powerful AI (GPT) for entity recognition and summarization, creating a synergistic approach that is both efficient and user-centric.
  • Structured Workflow for Complex Tasks: The "insight finding-scripting-storytelling" workflow provides a clear, step-by-step process for users, reducing cognitive load and making the complex task of storyline generation more manageable.
  • Interconnected Views for Seamless Interaction: The tight integration between the Text View, Fragment View, and Storyline View ensures that changes in one view are reflected in others, creating a fluid and intuitive user experience.

Key Results

  • Improved Storyline Extraction: Case studies and user interviews indicated that users could more effectively extract storylines using StoryExplorer compared to traditional methods.
  • High Content Similarity: When comparing user-generated storylines to a ground truth, an average content similarity of 90% for Task 1 and 85% for Task 2 was achieved.
  • Reduced Cognitive Load: Participants reported high intrinsic load (due to text complexity) but significantly low extraneous load (due to the tool), suggesting StoryExplorer is easy to learn and use.
  • Positive User Feedback: Usability questionnaire results showed favorable ratings for the workflow's comprehensibility, learnability, and usability, with participants enjoying the storyline creation process.
  • Efficiency Gains: System log analysis demonstrated efficiency in GPT-based entity annotation and a decrease in per-fragment operation time as users became more familiar with the system.

Limitations and Future Work

  • Temporal Juxtaposition: The current system relies on manual adjustments or text order for temporal placement. Future work could leverage more sophisticated temporal analysis of text and entity information for better automatic alignment.
  • Intelligent Fragment Recommendation: Users expressed a desire for more automatic fragment extraction. Future research could explore advanced event extraction algorithms to assist in this process.
  • Performance Bottlenecks: GPT model interface calls can lead to performance issues. Optimizations like text preprocessing and preloading entities are planned.
  • Limited User Demographics: The study participants were primarily students. Future evaluations should include a more diverse user base to assess applicability across different age groups and academic backgrounds.
  • Broader Text Type Applicability: While effective for novels and scripts, StoryExplorer's applicability to other text types like news articles and conference abstracts needs further investigation.
  • Long-Term Memory Impact: The current study focuses on the process of storyline generation. Future research will explore StoryExplorer's effectiveness in long-term memory retention through longitudinal studies.
  • Collaborative Features: The authors plan to explore enabling collaborative active reading and storyline construction through social platforms.
  • Neutral Evaluation Metrics: Future studies will use more neutral phrasing in questionnaires to avoid potential framing effects.

Summary updated

Abstract

In the context of the exponentially increasing volume of narrative texts such as novels and news, readers struggle to extract and consistently remember storylines from these intricate texts due to the constraints of human working memory and attention span. To tackle this issue, we propose a visualization approach StoryExplorer, which facilitates the process of knowledge externalization of narrative texts and further makes the form of mental models more coherent. Through the formative study and close collaboration with two domain experts, we identified key challenges for the extraction of the storyline. Guided by the distilled requirements, we then propose a set of workflow (i.e., insight finding-scripting-storytelling) to enable users to interactively generate fragments of narrative structures. We then propose a visualization system StoryExplorer that combines stroke annotation and GPT-based visual hints to quickly extract story fragments and interactively construct storylines. To evaluate the effectiveness and usefulness of StoryExplorer, we conducted two case studies and in-depth user interviews with 16 target users. The result shows that users can conveniently and effectively extract the storyline by using StoryExplorer along with the proposed workflow.
BibTeX
@article{ye2025storyexplorer,
  author = {Li Ye and Lei Wang and Shaolun Ruan and Heyu Wang and Yuwei Meng and Yigang Wang and Wei Chen and Zhiguang Zhou},
  title = {{StoryExplorer: A Visualization Framework for Storyline Generation of Textual Narratives}},
  journal = {IEEE Trans. Hum. Mach. Syst.},
  year = {2025},
  url = {https://openreview.net/forum?id=VA5v3kRQeo},
}
Figure