4
$\begingroup$

I am new to Doc2Vec. As I understand Doc2Vec group similar documents based on the context of their words.

I have a set of newspaper documents and I want to identify what are the main topics of the newspapers (group 'Politics' news documents to one group, 'Sports' news documents to another group etc.) based on their content. Thus, I am interested in knowing;

  1. What variant of Doc2Vec is more suitable for this (dbow, dm)?
  2. What are the real world applications of Doc2Vec algorithm?
$\endgroup$
3
$\begingroup$

I will answer your second question first, doc2vec and word2vec both are primarily good representations of text data that capture the semantics of words and documents. So whenever you are working with text data, you need a representation for it and that is what word2vec and doc2vec provides. Now think of any real world task on text data, like document similarity, using doc2vec you can find cosine similarity between two documents easily, now think of real world applications for it, finding duplicate question on a site like stack overflow, ranking candidate answers for a question answering model, features for text cladsification like sentiment analysis ( word2vec doesnt work well here, the context of good and bad are quite similar so it struggles to differentiate between positive and negative reviews). So these are just representations, you can apply them to any NLP and a lot of IR task.
To answer your first question, a model is not just task dependent but also data dependent. So you can read Mikolov's paper to find out how each model works for the baseline tasks but a good idea is to try both models on your data and extrinsically evaluate which algorithm performs better.

$\endgroup$
2
  • $\begingroup$ Thank you for your really great answer. I am quiet unsure about this application you mentioned ranking candidate answers for a question answering model, can we do ranking in doc2vec? Can you please further elaborate? $\endgroup$
    – Volka
    Oct 10 '17 at 0:22
  • 1
    $\begingroup$ If you make an IR based question answering model, and you find a document that contains your answer, you need to rank paragraphs/sentences to find out which contain your answers and a similarity metric can be used to rank paragraphs to the semantically processed form of your query $\endgroup$ Oct 10 '17 at 7:43

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.