AnchorTextVis: A Visual Analytics Approach for Fast Comparison of Text Embeddings

Area of Science:

• Information Visualization
• Natural Language Processing
• Machine Learning

Background:

• Visual comparison of text embeddings is essential for understanding semantic differences and evaluating embedding models.
• Current methods lack visual consistency and AI assistance, resulting in high cognitive load and inefficient exploration.
• Existing techniques struggle to preserve users' mental models during comparative analysis.

Purpose of the Study:

• To introduce AnchorTextVis, a novel visual analytics approach for comparing text embeddings.
• To improve visual consistency and accelerate the exploration of semantic differences between text embeddings.
• To leverage AI for enhanced analysis and summarization of text embedding comparisons.

Main Methods:

• Developed AnchorTextVis, integrating AnchorMap (a dynamic projection algorithm) and Joint t-SNE for visual consistency.
• Utilized large language models (LLMs) for AI-assisted comparison and summarization of text embeddings.
• Introduced quantitative metrics: Shared KNN and Coordinate distance for comparative analysis.
• Designed intuitive visualizations and interactive tools for comparing text clusters and individual texts.

Main Results:

• AnchorTextVis preserves visual consistency in comparative regions through comparable dimensionality reduction algorithms.
• LLMs accelerate the exploration process by providing AI-assisted comparisons and summaries.
• Quantitative metrics offer objective measures for comparing text embeddings.
• Case studies and expert feedback confirm the approach's effectiveness and usefulness.

Conclusions:

• AnchorTextVis significantly enhances the visual comparison of text embeddings by maintaining user mental maps.
• The integration of dynamic projection algorithms and LLMs offers a powerful tool for semantic analysis.
• The approach reduces cognitive load and time required for exploring and comparing text embedding models.