User reviews, which contain rich user preferences and item characteristics, can help to better predict user ratings in collaborative filtering (CF). However, one major issue is that reviews are usually posted after users have interacted with items, which limits the ways to use reviews and the ability to improve the performance of existing CF methods. To address this issue, this paper proposes a review construction-based CF method including: 1) a review construction network (RCN), which constructs vector representations of reviews before users interact with items; 2) a rating prediction network (RPN), which shares the intermediate layers with RCN to improve rating prediction performance; 3) several learning tricks, e.g., gradient reversal, to further boost the performance. Moreover, other CF methods, e.g., matrix factorization (MF), can be easily integrated into the proposed method to further boost the performance. Extensive studies on real-world datasets demonstrate that the proposed method integrated with MF significantly outperforms the state-of-the-art CF methods in recommendation accuracy.