Currently Fasttext is producing sentence vectors by taking average of the normalized word vectors of words in the sentence. Is it the best way to come up with the sentence vector.
Or whether using a tfidf weighting of words and then subtracting the first PCA component from it as discussed in this paper: https://openreview.net/pdf?id=SyK00v5xx will work better. And is there any such implementation already there within fasttext. If so where and how to use it through python.
Also, do I need to remove stop words from the sentence when I compute its fasttext vector?
Also in the python binding of fasttext how are to compute the sentence vector. There seem to be no syntax for that. Any comments.