My algorithm also use word2vec to make the abstract more accurate, but it will cost more time to run the algorithm. Using word2vec costs 40ms per article on the same traning data.