Neural Machine Translation (NMT) has secured significant traction recently owing to its intrinsic capability to generate translations of exceptional accuracy. However, the efficacy of NMT is highly dependent on the availability of training data, and the performance of the translation model tends to diminish significantly in the absence of large-scale and high-quality datasets. To mitigate this issue, particularly for languages with insufficient data, a word replacement-based data augmentation (DA) method is proposed in this paper. It initially performs part-of-speech tagging on the original data, followed by the utilization of a word replacement algorithm to generate new monolingual data. Subsequently, the corresponding translations are generated using the back-translation (BT) method, thus constructing pseudo-parallel data. Through a set of experiments, the results indicate that, in contrast to both the baseline model and current machine translation methods, the proposed word replacement-based method significantly improves translation quality.