Reviews have been commonly used to alleviate the sparsity problem in recommender systems, which has significantly improved the recommender performance. The review-based recommender systems can extract users features and items from review texts. The existing models such as D-Attn and NARRE employ convolutional neural networks and a coarse-grained attention mechanism to code reviews that have been embedded using the static word embedding, ignoring the long distance text information and lacks interpretability. To overcome these problems, this paper proposes the DNRDR (Dual-feature Neural Recommender with Dual-attention using Reviews) model, which can extract dual features of review text and can also enhance the interpretability using the word-level and review-level attention mechanisms. The proposed model is verified by experiments and compared with the state-of-the-art models. Besides, the dual-level attention mechanism can be visualized to improve interpretability.