Rating-based methods (e.g., collaborative filtering) in recommendation can explicitly model users and items from their rating patterns, nevertheless suffer from the natural data sparsity problem. In other hand, user-generated reviews can provide rich semantic information of user preference and item features, and can alleviate the sparsity problems of rating data. In fact, ratings and reviews are complementary and can be viewed as two different sides of users and items, hence fusing rating patterns and text reviews effectively has the potential to learn more accurate representations of users and items for recommendation. In this paper, we propose a hybrid neural recommendation model to learn the deep representations for users and items from both ratings and reviews. Our model contains three major components, i.e., a rating-based encoder to learn deep and explicit features from rating patterns of users and items, a review-based encoder to model users and items from text reviews, and the prediction module for recommendation according to the rating- and review-based representations of users and items. In addition, considering that different reviews have different informativeness for modelling users and items, we introduce a novel review-level attention mechanism incorporating with rating-based representation as query vector to select useful reviews. We conduct extensive experiments on several benchmark datasets and the experimental results demonstrate that our model can outperform the existing competitive baseline methods in recommendations.