The goal of tmsamples is to simulate term frequency matrices (DTMs or TCMs) based on parameters from real or simulated probabilistic topic models.
This corpus simulation is a core part of my dissertation research. The basic idea is: if you can simulate a corpus using the functional form of a topic model, and that corpus retains the gross statistical properties of human language, then we can use simulated corpora to derive rules and metrics to fit well specified topic models. Conversely, we can avoid pathologically misspecified models.
You can install the GitHub version of tmsamples with:
library(remotes)
remotes::install_github("tommyjones/tmsamples")This is a basic example which shows you how to solve a common problem:
library(tmsamples)
#> Loading required package: Matrix
Nk <- 4
Nd <- 50
Nv <- 1000
alpha <- rgamma(Nk, 0.5)
beta <- generate_zipf(vocab_size = Nv, magnitude = 500, zipf_par = 1.1)
pars <- sample_parameters(alpha, beta, Nd)
doc_lengths <- rpois(Nd, 50)
dtm <- sample_documents(
theta = pars$theta,
phi = pars$phi,
doc_lengths = doc_lengths,
threads = 2 ## threads controls parallel computation
)
#> ==================================================