Topic Space Visualization

Abstract

Topic modeling algorithms such as Latent Dirichlet Allocation (LDA) typically represent documents as a weighted combination of topics. Therefore, generalized barycentric coordinates are a natural fit for the visualization of a topic space. However, spatial positions in a planar barycentric coordinate system are ambiguous for more than three coordinates. Our glyphs for representing documents in combination with layout guidelines help to reduce the positional ambiguity. With an increasing number of documents, barycentric coordinate embeddings suffer from overplotting and visual clutter like other embeddings, possibly even more so since document po- sitions are fully independent of each other. Our experiments with jittering, aggregating glyphs, and grids show potential to reduce these problems for barycentric and other layouts.

Publications

Kiesel, D., Riehmann, P., Fan, F., Ajjour, Y., Wachsmuth, H., Stein. B., Froehlich, B. Improving Barycentric Embeddings of Topics Spaces (Poster) Proceedings IEEE VIS 2018, Berlin, Germany, October 2018
Preprint Poster

Ajjour, Y., Wachsmuth, H., Kiesel, D., Riehmann, P., Fan, F., Castiglia, G., Adejoh, R., Froehlich, B., Stein, B. Visualization of the Topic Space of Argument Search Results in args.me Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 60-65, Brussels, Belgium, November 2018.
Preprint