awaisathar / lda.js

LDA-Based Topic Modelling in Javascript

lda.js

LDA-Based Topic Modelling in Javascript

Topic modelling means detecting “abstract” topics from a collection of text documents. The most common text book technique to do that is using Latent Dirichlet Allocation. Simply put, LDA is a statistical algorithm which takes documents as input and produces a list of topics. One catch is that you have to tell it how many topics you want. There’s much more to it but since this is not a tutorial post, I will stop here. (If you are interested in how it works, read the references given on the wiki page.)

Here's a Javascript version of LDA, based on my no-longer-functioning earlier work. For testing, I use a subset of the SMS Spam Corpus available here (and thus take no responsibility of the inappropriateness of the text within :) ). Each topic is represented as a word cloud; the larger a word, the more weight it has in the topic. The source sentences are displayed again with a bar which shows the percentage distribution of topics for that sentence. Hovering on each area in the bar would show you the words in the topic. You can of course replace it with any other text, change the number of topics using the slider, and press the 'Analyse' button to see it work.

lda.js's Issues

Is it normal that the same word appears in different topics?

Is it normal that the same word appears in different topics? For example,

Topic 1: topic text data
Topic 2: modeling topic introduction

An out of the ordinary request

Hi,

I would like to make an unusual request.
I started with playing with the LDA on top of Wikipedia articles (or actually excerpts form Wikipedia articles).

I find that on the one hand, results are very sensitive to the choice of parameters (# topics, # top terms per topic,# iterations,# of documents and size of each document) and on the other hand, each run takes a very long time, so self-learning is very slow.
I would very much like to have some guidelines/logic/intuition/huristics to the choice of parameters.

What do you think a good medium would be to do that?
Thanks
Ilan