HINTs: Sensemaking on Large Collections of Documents With Hypergraph Visualization and INTelligent Agents.

Area of Science:

• Information Visualization
• Human-Computer Interaction
• Natural Language Processing

Background:

• Sensemaking on large document collections is challenging, with existing topic- and entity-based methods limited by NLP model capabilities.
• Poorly designed prompts and visualizations can mislead users and reduce system trustworthiness.
• Large Language Models (LLMs) offer enhanced accuracy and customizability for document analysis.

Purpose of the Study:

• To address limitations in document sensemaking by integrating user analysis tasks and visualization goals into prompt-based data extraction.
• To present HINTs, a visualization (VA) system that combines entity- and topic-based approaches for corpus analysis.
• To enhance the trustworthiness and effectiveness of sensemaking systems through Model Alignment.

Main Methods:

• Developed a three-stage visualization pipeline: prompt-based entity and topic extraction, hypergraph modeling and hierarchical clustering, and enhanced space-filling curve layout.
• Integrated an LLM-based intelligent chatbot agent to facilitate interactive exploration and sensemaking.
• Conducted case studies and a comparative user study to demonstrate generalizability and effectiveness.

Main Results:

• HINTs effectively supports sensemaking on large document collections by combining entity and topic perspectives.
• LLM-based intelligent agents can address many sensemaking challenges, but visual hints are still essential for accurate interpretation.
• The system's approach to prompt-based data extraction, guided by user tasks and visualization goals, enhances Model Alignment.

Conclusions:

• Intelligent agents and interactive visualizations can be synergistically combined to improve corpus analysis.
• Visual hints remain critical for ensuring the trustworthiness and interpretability of LLM-assisted sensemaking.
• Further research is needed to deeply integrate interactive visualization and LLMs for advanced corpus analysis.