How can I make use of large amounts of unlabelled in-domain text in a supervised hate speech detection project?

machine-learning svm bert-language-model unsupervised-learning supervised-learning