Hate speech is demonstrably aimed at social tension and violence. Detection becomes increasingly difficult as overlapping emotional feelings occur. However, there are still several unresolved issues with informal and indirect targeting of negative communication, including sarcasm, misrepresentation, and praise for the target’s or society’s immoral behavior. In this study, we proposed a method for instance selection based on attention network visualization. The goal is to categorize, modify, and expand the number of training instances. To this end, we first used the lexicons of hate speech and online forums to train the embedding using transfer learning. Then, we used synonym expansion to the semantic vectors. The active learning approach was used to train the task using the result-label pairs. The entropy-based selection and visualization techniques help select unlabeled text for each active learning cycle. The approach is improved, and the number of training instances is increased to improve the model’s accuracy. The active learning cycles are repeated until all unlabeled texts are converted to labeled text. The semantic embedding and lexicon expansion improve the model receiver operating characteristics (ROCs) from 0.89 to 0.91. The bidirectional LSTM with attention and active learning achieved 0.90 for precision–recall. The learned model can visualize the position-weighted terms to illustrate why hate speech is classified.