Uses R to control MALLET and generate models with different numbers of topics. A Log Likelihood value is extracted from each model and the vector of Log Likelihood values from all models generated is plotted and inspected to see which model has the highest, and thus what number of topics best suits the corpus. For a Windows machine.

R2MALLET-loop.r

# setup system enviroment for R and MALLET
MALLET_HOME <- "c:/mallet-2.0.7" # location of the bin directory
Sys.setenv("MALLET_HOME" = MALLET_HOME)
Sys.setenv(PATH = "c:/Program Files (x86)/Java/jre7/bin")

# configure variables and filenames for MALLET
## here using MALLET's built-in example data 

# set list of topic numbers to iterate over
seq <- seq(2, 100, 1)

# create list to store output of loop
loop <- vector("list", length(seq))

# define loop
for (i in seq) {

  # folder containing txt files for MALLET to work on
  importdir <- "C:\\mallet-2.0.7\\sample-data\\web\\en"
  # name of file for MALLET to train model on
  output <- paste("tutorial.mallet", i, sep = ".")
  # set number of topics for MALLET to use
  ntopics <- i
  # set optimisation interval for MALLET to use
  optint <-  10
  # set number of iterations per model
  iter <- 1000
  
  # set file names for output of model, extensions must be as shown
  outputstate <-     paste("topic-state", i, "gz", sep = ".")
  outputtopickeys <- paste("tutorial_keys", i, "txt", sep = ".")
  outputdoctopics <- paste("tutorial_composition", i, "txt", sep = ".")
  diagnostics <-     paste("diagnostics", i, "xml", sep = ".")
  
  # combine variables into strings ready for windows command line
  cd <- "cd C:\\mallet-2.0.7" # location of the bin directory
  import <- paste("bin\\mallet import-dir --input", importdir, "--output", output, "--keep-sequence --remove-stopwords", sep = " ")
  train  <- paste("bin\\mallet run cc.mallet.topics.tui.TopicTrainer --input", output, "--num-topics", ntopics, "--optimize-interval",  optint, "--output-state", outputstate, "--output-topic-keys", outputtopickeys, "--output-doc-topics", outputdoctopics, "--diagnostics-file", diagnostics, "--optimize-burn-in 200", sep = " ")
  
  # send commands to the Windows command prompt and run MALLET from there
  # collect console output in a list
  loop[[i]] <- shell(shQuote(paste(cd, import, train, sep = " && ")), 
                     invisible = FALSE, intern = TRUE)
}

# extract LogLiklihoods to see what number of topics has lowest LL
library(plyr)
LL1 <- ldply(seq, function(x) loop[x][[1]][[10]])
LL2 <- ldply(1:nrow(LL1), function (x) as.numeric(strsplit(LL1$V1, " ")[[x]][3]))
LLdf <- cbind(seq, LL2) 
# plot LL by number of topics
library(ggplot2)
library(grid)
ggplot(LLdf, aes(x = seq, y = V1)) + 
  xlab("Number of topics") + 
  ylab("Log likelihood of the model") + 
  geom_line() + 
  theme_bw()  + 
  theme(axis.title.x = element_text(vjust = -0.5, size = 14)) + 
  theme(axis.title.y=element_text(size = 14, angle=90, vjust= -0.25)) + 
  theme(plot.margin = unit(c(1,1,2,2), "lines"))

# inspect sorted table to find the number of topics with the lowest LL
# in case it's not easy to read it from the plot
LLdf[order(-LLdf$V1), ]