Constraint representation towards precise data-driven storytelling

Yu-Zhe Shi, Haotian Li, Lecheng Ruan, Huamin Qu

Abstract

Automating data-driven storytelling requires a delicate balance between top-down narrative intent and bottom-up data evidence. This paper proposes a framework of hierarchical constraints represented through Domain-Specific Languages (DSLs) to define a "permissible space" for storytelling. By formalizing the boundaries where a story remains both engaging and factually grounded, the framework aims to reduce narrative hallucinations and ensure logical consistency in automated systems.

Introduction

Creating a compelling data story is currently a manual, labor-intensive process. While modern AI has made significant strides in generating isolated text and charts, "data-driven storytelling" remains a complex hybrid challenge. Traditional storytelling is typically top-down, starting with a seed idea and rationalizing a world around it. Conversely, data reporting is bottom-up, summarizing evidence to reflect objective reality.

A successful data story must perform both functions simultaneously: it must convey a specific, often opinionated perspective while ensuring every claim is strictly supported by evidence. The primary obstacle to automation is the lack of a formal method to represent the constraints that keep these two forces in balance. Without such constraints, AI-generated stories often suffer from "hallucinated" narrative leaps or become bogged down in irrelevant data dumps.

A Hierarchical Framework of Constraints

The authors introduce a cognitively inspired framework that organizes storytelling constraints into two primary hierarchies: Interpretation and Articulation. Within these hierarchies, they identify ten specific types of constraints categorized across three levels of abstraction:

  1. Seed Idea to Articulation (The "Why" and "Who"): This level focuses on high-level narrative goals. It includes Theme-oriented constraints, which dictate the narrative structure, and Audience-oriented constraints, which govern the style and complexity of the output.
  2. Articulation to Interpretation (The "How"): This level ensures logical flow. It includes Inter-argument relational constraints (e.g., establishing contrast or analogy) and Spatial-temporal structural constraints to maintain consistency over the story's timeline.
  3. Evidence to Interpretation (The "What"): This level focuses on factual grounding. It involves Quantitative/Qualitative hypothesis spaces and Domain-specific knowledge to identify what constitutes a meaningful "insight" within a specific context.

Implementation via Domain-Specific Languages (DSLs)

To make these constraints machine-readable, the authors advocate for the use of Domain-Specific Languages (DSLs) rather than General-Purpose Languages (GPLs) like Python. DSLs provide a higher level of abstraction tailored specifically to storytelling, making it easier for domain experts—who may not be professional programmers—to verify that a generated story is complete, consistent, and grounded.

The paper outlines three levels of constraint logic within these DSLs:

  • Syntax-level: Manages the structural organization of arguments (e.g., trees or graphs).
  • Semantics-level: Ensures the meaning is consistent with common sense and domain-specific facts.
  • Execution-level: Dynamically verifies hypothesis spaces and variable lifecycles in real-time as the story is generated.

Exemplar Case Study: The Yellow River

As a theoretical viewpoint paper, the authors validate their framework through a detailed qualitative case study titled "The Protection of the Yellow River." By mapping this complex real-world narrative to formal DSL representations, they demonstrate the framework's utility:

  • Narrative Flow: Theme-oriented constraints were coded to ensure a logical "past-present-future" progression.
  • Insight Detection: Domain-specific knowledge constraints were used to flag significant data points—such as recognizing that 163 Fellows in a 300-person academy represents a critical mass—which standard statistical models might overlook.
  • The Permissible Space: The study illustrates how constraints define a bounded space for the AI, preventing it from veering into pure fiction or dry data summaries.

Conclusion and Future Work

The authors conclude that explicitly disentangling objective, subjective, and context-dependent constraints is a necessary step toward precise automated storytelling. However, several challenges remain. Currently, these DSLs must be manually designed by experts, which is a costly and time-consuming process.

Future research directions include the development of AutoDSL—using AI to automatically synthesize constraint languages from domain-specific corpora. Furthermore, the authors emphasize a "human-in-the-loop" approach, where constraints act as a bridge that allows human storytellers to focus on high-level intent while the system ensures technical grounding and consistency.

Summary updated

Abstract

Data-driven storytelling serves as a crucial bridge for communicating ideas in a persuasive way. However, the manual creation of data stories is a multifaceted, labor-intensive, and case-specific effort, limiting their broader application. As a result, automating the creation of data stories has emerged as a significant research thrust. Despite advances in Artificial Intelligence, the systematic generation of data stories remains challenging due to their hybrid nature: they must frame a perspective based on a seed idea in a top-down manner, similar to traditional storytelling, while coherently grounding insights of given evidence in a bottom-up fashion, akin to data analysis. These dual requirements necessitate precise constraints on the permissible space of a data story. In this viewpoint, we propose integrating constraints into the data story generation process. Defined upon the hierarchies of interpretation and articulation, constraints shape both narrations and illustrations to align with seed ideas and contextualized evidence. We identify the taxonomy and required functionalities of these constraints. Although constraints can be heterogeneous and latent, we explore the potential to represent them in a computation-friendly fashion via Domain-Specific Languages. We believe that leveraging constraints will facilitate both artistic and scientific aspects of data story generation.
BibTeX
@inproceedings{shi2024constraint,
  author = {Yu-Zhe Shi and Haotian Li and Lecheng Ruan and Huamin Qu},
  title = {{Constraint representation towards precise data-driven storytelling}},
  booktitle = {GEN4DS@VIS},
  year = {2024},
  url = {https://openreview.net/forum?id=hT2WH5OyRh},
}
Figure