Could you share how to prepare the word embeddings? Is it the pretrained word2vec, or some embeddings trained on the datasets like TREC4&5/Clueweb09?