Visualising Topics in Document Collections

Abstract

This paper discusses two multivariate visualisations which provide insights into topic model distributions across subcollections of a collection of historical textbooks in the context of a digital humanities project. Results of a qualitative user study with experts in historical research indicate that network graphs are more appropriate for revealing general connections among subcollections, while small-multiples of heatmaps of topic correlations are more suitable for a finer grained analysis of the connections between specific topics. We analyse the user behaviour during analysis to identify general activities of the interpretation of topic models as well as activities of interpreting visual elements that are specific to each visualisation. As a result, we observed usability problems and show potential for improved visualisations in digital humanities applications.