2012 Poster Sessions : Text Visualization Techniques for Assessing Topic Model Quality

Student Name : Jason Chuang
Advisor : Jeffrey M. Heer
Research Areas: Graphics/HCI
Topic models can help reveal patterns in large text corpora by identifying latent topics that capture recurring word usage. Successful real-world deployment of topic models, however, often requires intensive expert verification and significant model refinement. In this paper, we present Term Analyzer, a visual analysis tool for assessing topic model quality. We describe a distinctiveness measure for highlighting relevant terms, and introduce a novel seriation approach to surface clustering structures among related terms. We demonstrate that the tool allows analysts to more easily identify coherent themes and discard "junk topics."

Jason Chuang is a Ph.D. Candidate advised by Jeff Heer and Chris Manning. His research interests are in application of human-computer interaction techniques to advance the design of model-driven text visualizations.