An improved collaborative filtering algorithm was proposed aiming at the problems that user ratings for current e-commerce websites are too concentrated, the distinguishing degree is not obvious, and the credibility of the integer rating is not very well. Firstly, the improved part-of-speech path template algorithm was used to mine the product features and sentiment words contained in user’s reviews, and then the preference vector of the review was analyzed and established. Secondly, the emotional attitudes included in the review were calculated according to the review’s preference vector, and the emotional attitudes were used to revise the user’s rating, so that the revised rating was closer to the user’s true score willingness. Finally, the revised rating similarity and the preference similarity were combined to produce recommendations. The experimental results show that the proposed algorithm can effectively increase the classification and credibility of user’s ratings, and thus improves the quality of the nearest neighbors, as well as the accuracy of the recommended results.