Plume: Scaffolding Text Composition in Dashboards

Research Background and Issues

• What problems or challenges did the authors identify?
  Most data visualization and dashboard design tools focus on graphical representation, with very limited support for text creation. Crafting effective dashboard text is often considered a lengthy and unstructured process, especially when frequent text revisions are required due to data updates or varying presentation scenarios. Additionally, current tools and technologies lack structured and consistent methods to assist in text creation, particularly for the diverse functional roles of text in dashboards (e.g., labels, annotations, insights).

• Why is this issue important?
  Text in dashboards plays an indispensable role in providing context, insights, guiding interactions, and summarizing key information. The lack of tools that support text creation makes it difficult for dashboards to effectively convey data narratives, thereby impacting users' comprehension and decision-making capabilities.

• Research Motivation and Related Work
  The authors were inspired by prior work on the semantic roles and styles of dashboard text and identified the missing text creation functionalities in mainstream dashboard tools. They aim to address this technological gap through innovative tool design to better support dashboard authors in crafting efficient, clear, and readable text.

Solution

• What methods or solutions did the authors propose?
  The authors developed a system called Plume to assist in dashboard text creation. The system leverages large language models (LLMs) to provide context-aware text generation and incorporates a human-supervised workflow to help dashboard authors produce readable and semantically meaningful text. Additionally, the system introduces a "dashboard framework tree" to organize the visual layout and semantic relationships of the text.

• What are the innovative aspects of this solution?

  • Introducing semantic role-based classification (e.g., labels, insights, guidance) to generate dashboard text, marking the first structured support method for dashboard text.
  • Combining LLMs with human supervision to enable automatic text generation while allowing users to edit, simplify, or regenerate text as needed.
  • Incorporating a dashboard framework tree to semantically manage the hierarchical structure of the dashboard layout, ensuring that the generated text aligns with the visual structure.

• What are the implementation steps and key technologies used?

  1. Dashboard Framework Tree Definition: A tree structure is used to clearly define the scope of each text segment (e.g., targeting a single chart or the entire dashboard page).
  2. Text Generation and Suggestions: Plume generates semantically meaningful text based on predefined rules and dashboard content, with users able to make adjustments.
  3. Text Optimization and Feedback: The system provides suggestions for improving text and uses metrics like the Flesch-Kincaid readability score to assess text quality.
  4. User Interaction Design: Dynamic suggestions and lightweight user operations (e.g., accepting suggestions, modifying text) streamline the text creation process.

Research Outcomes

• What specific outcomes were achieved?

  • The system enables dashboard authors to generate and edit text at various levels, covering functional sections such as labels, insights, and contextual descriptions.
  • Preliminary user evaluations indicate that dashboard authors highly value the efficiency and accuracy of the generated text, while also expressing clear needs for flexibility and control options in the tool.

• What advantages does it have compared to existing solutions?

  • Plume begins with the classification of text roles, not only generating text but also providing assistance and feedback based on semantics and readability.
  • The dashboard framework tree enhances the association between text and visual elements, significantly improving user comprehension and navigation experience.
  • The introduction of a dynamic real-time update mechanism ensures that dashboard text aligns with changing data needs.

• What were the experimental or evaluation results?

  • In experiments with 12 dashboard authors, Plume achieved a system usability score (SUS) of 72.2, indicating a positive user experience.
  • On average, authors created 38 text segments, with approximately 66% directly accepted by users and about 25% optimized through user edits or tool regeneration.
  • Users highlighted the tool as a valuable starting point for drafting and iterating text, saving significant time.

• Limitations and Future Directions

  • Limitations: The generated text by Plume may sometimes be overly verbose or fail to fully capture the dashboard author's intent; dynamic updates may lead to content consistency or accuracy issues; cross-language support is still under development.
  • Future Directions:
    1. Explore long-term workflow integration to study how users gradually trust and adopt generated text in real-world tasks.
    2. Enhance support for text generation in complex chart types, potentially optimizing the creation process through specific data analysis models.
    3. Expand to collaborative workflows and multilingual support, such as enabling cross-language text generation and translation features.
    4. Develop more granular user control features to balance dynamic text updates with author verification needs.