8
$\begingroup$

I want to do document classification by representing each document as a set of features. I know that there are many ways: BOW, TFIDF, ...

I want to use Latent Dirichlet Allocation (LDA) to extract the topic keywords of EACH SINGLE document. the document is represented by these topic words. But I do not know whether it is reasonable because in my opinion LDA is usually used to extract the topic words shared by A BUNCH OF documents.

Can LDA be used to detect the topic of A SINGLE document?

$\endgroup$
  • $\begingroup$ For clarity, latent semantic allocation (LSA) or latent Dirichlet allocation (LDA)? $\endgroup$ – Sean Easter Sep 8 '14 at 20:55
  • $\begingroup$ @SeanEaster I mean "latent Dirichlet allocation" $\endgroup$ – Munichong Sep 10 '14 at 23:29
12
$\begingroup$

Can LDA be used to detect the topic of A SINGLE document?

Yes, in its particular representation of 'topic,' and given a training corpus of (usually related) documents.

LDA represents topics as distributions over words, and documents as distributions over topics. That is, one very purpose of LDA is to arrive at probabilistic representation of each document as a set of topics. For example, the LDA implementation in gensim can return this representation for any given document.

But this depends on the other documents in the corpus: Any given document will have a different representation if analyzed as part of a different corpus.

That's not typically considered a shortcoming: Most applications of LDA focus on related documents. The paper introducing LDA applies it to two corpora, one of Associated Press articles and one of scientific article abstracts. Edwin Chen's nicely approachable blog post applies LDA to a tranche of emails from Sarah Palin's time as Alaska governor.

If your application demands separating documents into known, mutually exclusive classes, then LDA-derived topics can be used as features for classification. Indeed, the initial paper does just that with the AP corpus, with good results.

Relatedly, Chen's demonstration doesn't sort documents into exclusive classes, but his documents' mostly concentrate their probability on single LDA topics. As David Blei explains in this video lecture, the Dirichlet priors can be chosen to favor sparsity. More simply, "a document is penalized for using many topics," as his slides put it. This seems the closest LDA can get to a single, unsupervised topic, but certainly doesn't guarantee every document will be represented as such.

$\endgroup$
  • $\begingroup$ Thanks for you reply. Can I input one single document, rather than a set of documents, into LDA and use the output words as the topic of the document? $\endgroup$ – Munichong Sep 12 '14 at 0:20
  • $\begingroup$ By "input one single document," do you mean "use a training set consisting of a single document"? Or, "extract topics from a single, out-of-sample document using a trained model"? $\endgroup$ – Sean Easter Sep 12 '14 at 13:30
  • $\begingroup$ I mean "use a training set consisting of a single document" and specify the number of target topics to 1, that is, extract one topic (a set of words) from the only one training document. $\endgroup$ – Munichong Sep 12 '14 at 13:47
  • $\begingroup$ I've never tried, so I can't positively confirm how any implementation would treat that input. But using a single document and single topic would defy certain generative assumptions of the model. You'd essentially be assuming each word in the document was drawn from a single multinomial distribution; the posterior for its parameter would be little more than a normalized count of words in the document. All this to say I struggle to see how this could produce meaningful output. It'd be like calculating tf-idf on a corpus of one: The statistic relies on characteristics of the broader corpus. $\endgroup$ – Sean Easter Sep 13 '14 at 16:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.