Large-Scale Evaluation of Topic Models and Dimensionality Reduction Methods for 2D Text Spatialization.

Area of Science:

• Computational linguistics
• Data visualization
• Machine learning

Background:

• Topic models uncover semantic structures in text corpora.
• Topic models combined with dimensionality reduction create document spatializations.
• Optimal combinations for high-quality text spatializations are unknown.

Purpose of the Study:

• Evaluate topic models and dimensionality reduction for text spatialization.
• Identify effective combinations for accurate and perceptually effective layouts.
• Provide guidelines for designing text visualizations.

Main Methods:

• Large-scale computational evaluation using a benchmark dataset.
• Tested combinations of topic models and dimensionality reduction algorithms.
• Quantified layout quality using metrics for local/global properties and perceptual effectiveness.

Main Results:

• Generated over 45,000 unique layouts and quality metrics.
• Interpretable topic models significantly improve corpus structure capture.
• t-Distributed Stochastic Neighbor Embedding (t-SNE) is recommended for dimensionality reduction.

Conclusions:

• Guidelines for effective text spatialization design are proposed.
• Interpretable topic models are crucial for semantic structure representation.
• t-SNE is a highly effective dimensionality reduction technique for this task.