After training word vectors with word2vec, is it better to normalize them before using them for some downstream applications? I.e what are the pros/cons of normalizing them?