Entity resolution is an important task in data integration and data cleaning. Two major problems in this context are the availability of labeled data and the quality of these labels. When few data points are available, a frequently used approach is data augmentation. Current data augmentation techniques yield inconsistent and small increases in performance on this task, especially when combined with transformer-based language models. This study proposes a graph-based approach that makes use of positive and negative labels to represent the data. By recombining entities, this allows generating new data points without introducing additional label noise. Furthermore, it also provides two ways of identifying existing label noise through consistency checks and finding loss-outliers in the graph-structure. Experiments on ten datasets using two distinct transformer-based models are conducted, comparing the performance of existing methods to the proposed method. These experiments show an improvement in F1 score of up to 6% with a median improvement of 3%, tripling the performance gains from existing approaches. On multiple data sets this proposed method effectively provides the same increase in performance as labeling more than three times the amount of data.