GSPbtm provides an easy application of the BTM package based on the paper.

Great for use with any short text dataset with three types of models - NAV (noun, adjective, verb); NPN (noun, proper noun); (ADJ) (adjective only). There is a single wrapper function complete_btm() that will run the full model. User specifies the minimum and maximum number of topics to estimate and the type of model to run, and the model will select the best one based upon a calculated coherence score. Because we use a background topic supplied by BTM, the topics are generally well defined and informative; coherence scores are not overwhelmed by common stop words or uninformative terms.

The column with the text to be annotated and modeled must be named “text” and the user will be prompted if no column has the name. Annotation relies upon the udpipe package to POS tag and determine biterm coocurrence.

Once the best topic is selected, we borrow the Jensen-Shannon Divergence and PCA from the LDAvis package, and extract our X,Y coordinates for each topic. The top 12 terms are selected using the relevance and frequency selection criteria from the LDAvis package and returned as a separate output (note that the relevance/frequency split is hardcoded at 0.6).

You can install the development version of GSPbtm from GitHub with:

# install.packages("devtools")
devtools::install_github("taylorgrant/GSPbtm")

Assuming a data frame with a column named “text”:

library(GSPbtm)
tm_out <- complete_btm(data, min_topics = 5, max_topics = 20, model = "NAV")

Function returns a list: tm_out$df is the original dataframe with topic numbers joined per row; tm_out$fulldata - data of top n terms with topic numbers and PCA coordinates; tm_out$range - the min and max to set axes when plotting.