Sensecape: Enabling Multilevel Exploration and Sensemaking with Large Language Models

Document Title

Sensecape: Enabling Multilevel Exploration and Sensemaking with Large Language Models

Document Information

• Domain: Information exploration, application of AI-generated language models in complex information tasks
• Keywords: Information seeking, multilevel exploration, information organization and understanding, abstraction levels, human-computer interaction, large language models, systems thinking, visual spatial information diagrams

Research Background and Problems

• Challenges or Issues:

  • Current user interfaces based on large language models (LLMs) primarily rely on linear conversational interactions, whereas most complex information tasks (e.g., academic research, project planning) typically require nonlinear methods of information organization and summarization.
  • In complex information tasks, users need to switch between information retrieval and sensemaking, but existing user interfaces often fail to support such workflows effectively.
  • The large volume of information generated by LLMs can lead to cognitive overload and visual clutter, while linear conversation records make information retrieval more difficult.

• Research Motivation:

  • Information exploration and sensemaking are central to complex cognitive tasks, and leveraging the powerful generative capabilities of LLMs to optimize this process remains an unresolved challenge.
  • The authors aim to improve the interaction between LLMs and users, enabling users to handle information tasks in a more structured and efficient manner.

• Related Work:

  • Research in information retrieval (IR) and human-computer interaction (HCI) focuses on exploratory search, sensemaking, and tools that support information organization and user cognitive structures. For instance, existing tools like InkSeine and Miro offer some exploratory features but still fall short in organizing and integrating information hierarchically.

Solution

• Proposed Solution:

  • The authors designed an interactive system called Sensecape, which supports users in multilevel exploration and sensemaking of information tasks using large language models.
  • The Sensecape system integrates two core views: the Canvas View for nonlinear information presentation and the Hierarchy View for multilevel abstraction navigation, enabling users to seamlessly switch between different abstraction levels while managing and organizing information.

• Innovations:

  • Multilevel abstraction support: By visualizing abstraction levels, the system helps users develop a clearer understanding of the information space.
  • Integration of intelligent features such as Semantic Zoom and Semantic Dive, allowing users to adjust the granularity of information or delve deeper into specific topics as needed.
  • Explicitly presenting the organizational structure of information through the hierarchy view, encouraging users to build more systematic knowledge layers.

• Implementation Steps and Key Technologies:

  1. Provide an information canvas feature that allows users to freely drag, group, connect, and edit nodes.
  2. Enable dynamic navigation through abstraction levels, supporting users in switching between high-level overviews and detailed insights.
  3. Integrate multiple exploration functionalities (e.g., generating subtopics, generating related questions) and semantic zoom technology to adjust content detail levels.
  4. The system implementation utilizes the React framework and integrates OpenAI's GPT-4 model to generate task-relevant content.

Research Outcomes

• Specific Results:

  • Experimental results demonstrate that the Sensecape system significantly enhances users' ability to explore and organize information hierarchically in complex topics.
  • User testing showed that Sensecape enables users to explore multilevel information more efficiently and promotes better integration of information.

• Comparison with Existing Solutions:

  • Compared to existing tools (e.g., combining ChatGPT with Miro), Sensecape places greater emphasis on unifying exploration and organization within the information workflow, allowing users to complete tasks more efficiently.
  • Compared to tools without hierarchy views, Sensecape helps reduce visual clutter and guides users in abstracting complex information through semantic zoom and hierarchical structuring.

• Experimental or Evaluation Results:

  • Users significantly increased the number of concepts explored in the Sensecape environment (from a baseline of 22.8 to 68.3) and were able to create deeper knowledge hierarchies.
  • Participants generally found the hierarchy view and semantic zoom particularly helpful in addressing cognitive overload issues.
  • Although some participants mentioned a steep learning curve, most provided positive feedback.

• Limitations and Future Directions:

  • Limitations:
    • The interface complexity posed cognitive challenges for some users.
    • The short duration of the experiment may have limited users' full understanding of the system's capabilities.
  • Future Directions:
    • Enhance Sensecape's collaboration features to support team-based knowledge construction.
    • Improve visual and interaction design to reduce user learning costs.
    • Consider adding real-time context recommendations and information verification features to increase the tool's practical value.
    • Explore the application of multilevel exploration in other abstract domains (e.g., software development, system design).