Applying convolutional neural network over text documents using 1-D tf-idf feature vectors

Question

I want to apply a CNN over documents. I have tf-idf vectors of documents with me (one vector per document).

My question is, is 1D CNN applicable in this case? The reason I am asking this question is that I have read 1D convolutions being applied to sequence of vectors and I have sequence of scalars (tf-idf vector) as input in this case, so is it applicable?

gergelybat · Answer

In TFIDF representation "neighborhood" is not interpretable. The 1st number in a document vector has just as much relation to the 2nd and to the 100th.
While in case of a standard time series or sequence analysis "neighborhood" is interpretable.
On the other hand, if you could re-order the TFIDF representation in a way, that makes sense from the analysis perspective (eg: ordering based on the "sentiment" of the word), maybe 1D CNN would show some good results.

Applying convolutional neural network over text documents using 1-D tf-idf feature vectors

One Answer

Add your own answers!

Ask a Question