Due to the extensive use of social media platforms for several purposes like expressing several forms of hatred, researchers have experimented with many machine learning and natural language processing models to identify online hate speech automatically. Big proportion of these models depend heavily on labeled datasets to be trained. However, labeled datasets, especially multi-label ones, suffer from certain challenges such as imbalanced label distributions and restricted sizes. This research investigates the efficiency of using and combining different data-augmentation methods to address the problem of limited size of data and to enhance its quality. Four distinct augmentation approaches were used: synonym replacement, token replacement, paraphrasing, and back translation. These four were combined and evaluated, and the findings indicate that combining significantly improves performance by recording the macro F1 of 0.8311 on the ETHOS multi-label dataset.