Law articles are the main basis of legal judgment and related law article recommendation can improve the quality of legal judgment prediction (LJP). Currently, the state-of-the-art methods belong to supervised classification model with the finite law articles as discrete class labels, which, however, has the downside that the semantic information of law articles may be underused. This observation implies that the quality of law article recommendation can be further enhanced. To solve this problem, this study proposes a deep semantic matching based law article recommendation approach (DeepLawRec) by transforming the traditional classification solution to pairwise matching learning. The proposed DeepLawRec includes local semantic matching module and global semantic recommendation module with a bi-transformer convolutional neural network and regression tree based recommendation, respectively. It can not only extract the key semantic features from the fact descriptions of cases, but also learn their related and local semantic features with a given law article. Moreover, with the help of regression tree, the recommendation results could be interpreted. The experimental results on a public dataset show the proposed DeepLawRec approach can improve the quality of recommendation and outperform the state-of-the-art techniques in terms of precision and interpretability.