The prospective notification on tweet streams is a challenge task in which the user wishes to receive timely, relevant, and non-redundant update notification to remain up-to-date. To be effective the system attempts to optimize the aforementioned properties (timeliness, relevance, novelty and redundancy) and find a trade-off between pushing too many and pushing too few tweets. We propose an adaptation of the extended Boolean model based on word similarity to estimate the relevance score of tweets. We take advantage of the word2vec model to capture the similarity between query terms and tweet terms. Experiments on the TREC MB RTF 2015 dataset show that our approach outperforms all considered baselines.