Topic Scaling: A Joint Document Scaling–Topic Model Approach to Learn Time-Specific Topics

This paper proposes a new methodology to study sequential corpora by implementing two-stage algorithm that learns time-based topics with respect scale of document positions and introduces the concept Topic Scaling, which ranks learned within same scale. The first stage documents using Wordfish, Poisson-based document-scaling method, estimate serve, in second stage, as dependent variable learn relevant via supervised Latent Dirichlet Allocation. novelty brings two innovations text mining it explains positions, whose is latent variable, inferred on match their occurrences corpus track evolution. Tested U.S. State Of Union two-party addresses, this inductive approach reveals each party dominates one end interchangeable transitions follow parties’ term office, while shows for German economic forecasting reports shift narrative style adopted institutions following 2008 financial crisis. Besides demonstrated high accuracy predicting in-sample from topic scores, method unfolds further hidden differentiate similar increasing number expand potential nested hierarchical structures. Compared other popular models, Scaling similarities without specifying time frequency evolution, thus capturing broader patterns than dynamic models yielding more interpretable outputs plain