This package implements an approach to quickly fitting topic models, combining partial PCA for sparse matrices with a varimax rotation, proposed by Rohe and Zang (https://arxiv.org/abs/2004.05387). In simulation, as implemented here this method runs roughly an order of magnitude faster than structural topic models from the stm package. The method is also deterministic and does not introduce research degrees of freedom through the Bayesian priors of LDA.
Beyond fitting the topic models, the package includes (a) functions for my information-theoretic approach to vocabulary selection; (b) tidiers, for extracting both word-topic and topic-document matrices into a tidyverse workflow; (c) Hellinger distance calculations and t-SNE and UMAP visualization for my "discursive space" analysis; and (d) samplers to construct simulated corpora.
A preprint discussing the package is available on the arXiv.
remotes::install_github("dhicks/tmfast")
or fork https://github.com/dhicks/tmfast, clone, and install manually.