An increasing number of online news triggers wide academic concern for the prediction of news popularity, which is affected by users' behaviors and not easy to predict. However, existing methods that predict the popularity of online news after publication are not timely enough, and predicting before publication lacks discriminatory features. This paper explores the variables which may affect news popularity and presents a novel methodology to predict the popularity of online news before publication. Through the observation of news, we first find that grammatical construction of titles can affect news popularity, and experiments show that this feature can improve R^2 statistics of the prediction model by 6.62% exactly. Besides, we improve traditional category and author features by using logarithmic conversion to views first and calculating a score of these features instead of stuffing them into learning models directly. Using these features and two other features, we finally predict news popularity in two aspects: whether the news will be popular and how many views the news ultimately attract.