Some researches had succeeded in doing hate speech identification automatically from text with machine learning and deep learning approaches. However, it was still unclear how adaptive is a deep learning-based model if it is tested on a different set of text data with different domain. To address this issue, this research proposed some deep learning-based methods, using some variants of Recurrent Neural Network to identify hate speech in texts sourced from Twitter, and then used to predict other set of text data sourced from Facebook and Twitter. The experiment was done in order to measure the difference of model performance between training phase and testing phase. Experiment results showed that the proposed method outperformed the machine learning based methods, both in training phase, by GRU algorithm with 85.37% F1-score, and in testing phase, by LSTM algorithm with 76.30% F1-score. Then, in terms of adaptability of model performance, the proposed method gave comparable result against the baseline method.