Deep Learning Based Hate Speech Detection on Twitter

There have been growing worries about the effects of the widespread use of hate speech and harsh language on social media sites like Twitter. Effective strategies for recognising and reducing such dangerous material are necessary for resolving this problem. In this research, we give a detailed analysis of four deep learning models for identifying hate speech and inflammatory language on Twitter: the Long Short-Term Memory (LSTM), the Recurrent Neural Network (RNN), the Bidirectional LSTM (Bi-LSTM), and the Gated Recurrent Unit (GRU). We downloaded a large dataset from Kaggle that was curated for hate speech identification and used it in our experiment. We built each model after preprocessing and tokenization, then tweaked their hyperparameters for maximum efficiency. The models' abilities to detect hate speech were evaluated using standard measures including accuracy, precision, recall, and Fl-score. Our findings show that there is a wide range of effectiveness amongst models in terms of identifying hate speech and inflammatory language on Twitter. In terms of accuracy and Fl-scores, the Bi-LSTM and GRU models were superior to the LSTM and RNN. The results of this study imply that using bidirectional and gated processes may increase the models' capability of understanding the interdependencies and contexts of tweets, and hence, their classification accuracy.