Telling Stories from Computational Notebooks: AI-Assisted Presentation Slides Creation for Presenting Data Science Work

Document Title

Telling Stories from Computational Notebooks: AI-Assisted Presentation Slides Creation for Presenting Data Science Work

Document Information

• Domain: Human-Computer Interaction and Data Science
• Keywords: Data science presentation, AI-assisted tools, computational notebooks, slide generation, human-machine collaboration, automation, human-centered design, deep learning

Research Background and Problem Statement

• Problems and Challenges:

  • After completing technical tasks, data scientists often spend significant time and effort creating presentation slides to explain models and results to stakeholders, which is both tedious and time-consuming.
  • Existing automation technologies (e.g., AutoML) primarily focus on technical aspects (e.g., model selection, feature engineering) and provide limited support for communication tasks such as slide generation.
  • Data science teams consist of diverse roles, and communication between technical and non-technical teams faces significant challenges, including knowledge gaps, trust-building, and expectation management.

• Significance:

  • Effective presentations and storytelling capabilities enable data scientists to help project stakeholders better understand complex technical work, ultimately facilitating feedback or decision-making.
  • Reducing the tedious process of creating presentation slides allows data scientists to focus their efforts on higher-value tasks.

• Research Motivation and Related Work:

  • There is a need for new tools that support human-machine collaboration, enabling data scientists to quickly generate presentation slides from computational notebooks while offering editable and customizable interactive features.
  • Existing research has explored areas such as code readability and automated documentation generation, but studies related to slide generation and presentation remain scarce.

Solution

• Method and Approach:

  • A human-centered AI system named "NB2Slides" is proposed to generate presentation slides from Jupyter notebooks.
  • The system combines deep learning models (e.g., code summary generation) with rule-based template methods (e.g., generating slides with different templates based on user configuration).
  • The slide generation process balances simplicity and flexibility, allowing users to quickly adjust content based on the target audience (technical or non-technical).

• Innovations:

  • Dynamically adjusts generated content based on the audience background provided by the user.
  • Links the generated slides to the original notebook content, offering traceability and explainability.
  • Integrates automation (AI model generation) with manual editing capabilities, emphasizing human-machine collaboration rather than complete automation.

• Implementation Steps and Key Technologies:

  1. Parse Jupyter notebook content and construct a hierarchical structure.
  2. Extract relevant code blocks from the notebook based on automatically generated templates.
  3. Use deep neural networks (e.g., T5 model) to generate summary content.
  4. Display the generated slides and provide interactive links to the notebook.
  5. Offer user editing tools to further refine the generated content.

Research Outcomes

• Specific Results:

  • Developed and tested the NB2Slides system, which can transform a completed Jupyter notebook into preliminary presentation slides.
  • The system dynamically generates content layouts based on the target audience type (e.g., hiding technical details for non-technical audiences while including more code details for technical audiences).

• Advantages:

  • Significantly improves the efficiency of data scientists in creating presentation slides.
  • Reduces the complexity of information organization and provides visual explanations of the generated content.
  • Supports partial automation and personalized content customization during the presentation process.

• Experiments and Evaluation Results:

  • A user study involving 12 experienced data scientists demonstrated that NB2Slides enhanced users' work efficiency, saving time in content organization and slide creation.
  • Users found the system-generated content generally accurate and well-structured but emphasized that complete automation still requires human intervention to supplement business context and technical storytelling.

• Limitations and Future Directions:

  • Relies on high-quality input notebooks and has limited ability to handle disorganized or partially incomplete notebook content.
  • The logical coherence and readability of generated content need further improvement, such as better code summarization models.
  • Future work may explore more refined generation configurations, such as supporting presentation needs for different scenarios and interactive content displays.