Slide4N: Creating Presentation Slides from Computational Notebooks with Human-AI Collaboration

Document Title

Slide4N: Creating Presentation Slides from Computational Notebooks with Human-AI Collaboration

Document Information

• Domain: Human-AI Collaboration and Data Science Visualization Tool Design
• Keywords: Slide Generation, Computational Notebooks, Human-AI Collaboration, Natural Language Processing, Data Science

Research Background and Problem

• Problem or Challenge:

  • Data scientists often need to manually extract content from computational notebooks (e.g., Jupyter Notebooks) to create presentation slides for both technical and non-technical audiences.
  • Notebook files are typically lengthy and disorganized, containing intermediate data and notes that are unsuitable for direct presentation as slides.
  • Frequent updates and manual operations make slide creation time-consuming and prone to errors.

• Significance:

  • Collaboration and communication are critical in data science, especially between technical and non-technical stakeholders. Simplifying the process from analysis to presentation is essential for improving efficiency and reducing repetitive tasks.

• Research Motivation and Related Work:

  • Existing solutions such as RISE, ToonNote, and NB2Slides have limitations:
    • Some tools merely convert raw notebook content without generating readable summaries or presentation-ready content.
    • Most tools impose strict requirements on notebook input quality and lack user interaction support.
  • This study proposes a human-AI collaborative approach and the Slide4N system, combining NLP techniques and user interaction design to balance automation and manual adjustments.

Solution

• Method or Solution:

  • Design Slide4N, an interactive AI assistant embedded in the JupyterLab environment, to automatically generate slides from user-selected notebook cells.
  • Leverage advanced natural language processing techniques to generate slide titles and key point summaries, analyze relationships between notebook cells, and design appropriate slide layouts.

• Innovations:

  • Introduces a human-AI collaborative slide generation method, avoiding the limitations of fully automated systems while allowing user intervention in the generation process.
  • Enables users to select specific notebook cells for content generation instead of relying on the entire notebook.
  • Binds presentation content to notebook content, ensuring synchronization between analysis and presentation.

• Implementation Steps and Key Technologies:

  • Identifying Relevant Cells: Build a code cell relationship graph, calculate relevance scores based on shared variables, and automatically suggest potentially related cells.
  • Information Extraction and Summarization:
    • Use the CodeT5 model to generate slide key points from code cells.
    • Employ HAConvGNN and T5 models to generate title candidates.
  • Content Organization and Layout: Apply hierarchical clustering algorithms and tree structures to design logical relationships and form slide group layouts based on content associations.
  • User Customization: Provide intuitive interaction support, allowing users to adjust titles, summaries, and slide layouts.

Research Outcomes

• Specific Outcomes:

  • Developed the Slide4N system, which can automatically generate slides and supports further user adjustments.
  • Compared to manual or fully automated methods, Slide4N is more efficient and produces higher-quality slides.

• Advantages:

  • Reduces the time required for slide creation and minimizes the effort of converting notebooks into presentation tools.
  • Enhances communication efficiency within collaborative teams, particularly between technical and non-technical members.

• Experimental or Evaluation Results:

  • User studies indicate high satisfaction with Slide4N, especially in identifying relevant cells and generating readable content summaries.
  • Slide content consistency and logical layout were well-received, though aesthetic quality was slightly lacking.
  • Slide4N supports varying levels of user intervention, meeting personalized needs.

• Limitations and Future Directions:

  • Slide4N's generation is currently limited to code cells, overlooking finer-grained content such as markdown text or code outputs.
  • The system has room for improvement in visualization support and slide style customization, such as enhanced editing capabilities and more export formats.
  • Future work could include long-term deployment studies, comparative studies with other methods, and improved privacy features, such as support for local execution.