MATLAB: Does the text analytics toolbox allow users to test out-of-sample perplexity with LDA

ldaText Analytics Toolbox

I want to create two samples from my data: one for training and one for testing. Then I want to fit the LDA model using the training sample. Then I want to test the preplexity of the test sample using the fitted model. Is this possible with the text analytics toolbox?

Best Answer

The second output of logp gives you the perplexity.
txt = extractFileText('sonnets.txt');
sonnets = split(txt,[newline newline]);
sonnets = sonnets(5:2:end);
td = tokenizedDocument(sonnets);
bow = bagOfWords(td(1:50));
mdl = fitlda(bow,5,'Verbose',0);
[~,perpl] = logp(mdl, encode(bow,td(51:53)))
% perpl = 337.4999
Related Question