Visualizing Temporal-Thematic Patterns in Text Collections

Abstract

Visualizing the temporal evolution of texts is relevant for many domains that seek to gain insight from text repositories. However, existing visualization methods for text collections do not show fine-grained temporal-thematic patterns. Therefore, we developed and analyzed a new visualization method that aims at uncovering such patterns. Specifically, we project texts to one dimension, which allows positioning texts in a 2D diagram of projection space and time. For projection, we employed two manifold learning algorithms: the self-organizing map (SOM) and UMAP. To assess the utility of our method, we experimented with real-world datasets and discuss the resulting visualizations. We find our method facilitates relating patterns and extracting associated texts beyond what is possible with previous techniques. We also conducted interviews with historians to show that our prototypical system supports domain experts in their analysis tasks.