Abstract: To solve the effect problem of sentiment classification due to the insufficient use of emotional semantic features and unpromising dimension reduction effects, a novel high-precision sentiment classification method was proposed in this paper for online comments by extending semantic similar emotional words and employing the statistical features between words. Firstly, a neural network skip-gram model was employed to generate word embedding and extend the semantic similar words to emotional feature by the measure of embedding word similarity. Then the feature dimension was reduced by employing the statistical features between words. At last, sentiment classification of online comments was carried out by the Adaboost classification model which was constructed by weighting multiple weak classifiers. Experiment results on hotel reviews and mobile comments show that, the accuracy of sentiment classification with new method can reach 90.96% and 93.67% respectively. Expanding semantic similarity emotion words is helpful to enrich the semantic features of emotion. Employing statistical features between words has better feature reduction effect. Both two procedures effectively improve the performance of text sentiment classification.