The goal of cross-language entity alignment is to connect the same entities in the cross-language knowledge graph to help downstream applications that need cross-language knowledge. At present, the main methods to realize cross-language entity alignment are knowledge graph representation learning method and graph convolution network method. The former uses the knowledge graph representation learning method to learn the embedded vectors of knowledge in different knowledge graphs and aligns entities by calculating the similarity between vectors, but this method cannot well capture the complex relationships commonly existing in multi-relational knowledge graphs. In contrast, the latter shows obvious advantages. However, pre-aligned entity pairs from manual annotation usually contain errors, which will affect the final alignment results. Therefore, this paper proposes a noise-sensitive relationship-aware dual graph convolution network (NSRDGCN), which is composed of noise detection and relationship-aware entity alignment. The noise detection part uses GAN to detect the noise in the training set, and the relationship-aware entity alignment part learns a better entity representation through the influence between the triples of (entity, relationship, entity) and its corresponding new triples (relationship, entity, relationship). Experiments on three cross language entity data sets show that our method is significantly better than the existing entity alignment methods. In particular, based on the benchmark data DBP15K \(_{\text {JA-EN}}\) , the entity alignment model NSRDGCN achieves a relative improvement of about 1.6% for Hits@10 compared to the RDGCN model.