Composing Data Stories with Meta Relations

TL;DR

This work addresses the gap in AI-powered data storytelling where data-level relations fail to capture domain knowledge and narrative intent. It formalizes meta relations as $relation_{AB} := (fact_A, fact_B, type, score)$ and develops Remex, a notebook-based tool that uses LLMs and rule-based verification to infer meta relations and organize data facts into slide-based stories. Through a formative study and an exploratory user study, the authors demonstrate that meta relations improve contextual understanding, storytelling direction, and cohesion between pieces, while highlighting challenges in domain knowledge acquisition, user workload, and AI reliability. The findings yield design considerations and a practical prototype that informs future human-AI collaboration in data storytelling, with potential extensions to broader formats and retrieval-augmented approaches.

Abstract

To facilitate the creation of compelling and engaging data stories, AI-powered tools have been introduced to automate the three stages in the workflow: analyzing data, organizing findings, and creating visuals. However, these tools rely on data-level information to derive inflexible relations between findings. Therefore, they often create one-size-fits-all data stories. Differently, our formative study reveals that humans heavily rely on meta relations between these findings from diverse domain knowledge and narrative intent, going beyond datasets, to compose their findings into stylized data stories. Such a gap indicates the importance of introducing meta relations to elevate AI-created stories to a satisfactory level. Though necessary, it is still unclear where and how AI should be involved in working with humans on meta relations. To answer the question, we conducted an exploratory user study with Remex, an AI-powered data storytelling tool that suggests meta relations in the analysis stage and applies meta relations for data story organization. The user study reveals various findings about introducing AI for meta relations into the storytelling workflow, such as the benefit of considering meta relations and their diverse expected usage scenarios. Finally, the paper concludes with lessons and suggestions about applying meta relations to compose data stories to hopefully inspire future research.

Figures (5)

• Figure 1: A motivating example to showcase the necessity of meta relations in data storytelling from Bloomberg pearce2015car.
• Figure 2: This figure illustrates the structure of Remex. More details about the analysis and the organization panels are available in Fig. \ref{['fig:interface']}.
• Figure 3: The figure explains the meta relation identification module.
• Figure 4: The figure explains the data story organization module.
• Figure 5: This figure shows Remex in JupyterLab with a case in the user study. Remex consists of multiple analysis panels under code cells and an organization panel to show how the data findings are organized into a sequence of slides. (a)-(c) show previous facts, meta relations, and current facts in both analysis and organization panels. The arrows with (a)-(c) indicate that two linked areas show the same information. (d1) and (d2) explain a slide and a fact in the slide in the organization panel. (d3) shows exported slides. (e1)-(e7) explain the interactions. The arrow cursor indicates a hover interaction, while the hand cursor indicates a click interaction.