An Inter and Intra Transformer for Hate Speech Detection

Hate speech is prevalent in the Internet environment, under circumstances of anonymity and physical distance on the Web. Hate speech detection has also had an increasing trend of development. In recent years, successful hate speech detection models have been developed, such as using CNN, LSTM, but they lack a sufficient amount of discriminative power. In order to tackle the task of hate speech detection, this paper introduces the use of center loss – known for its applications in image understanding – on top of the basis of the Transformer architecture. Our model can become more discriminative with the help of the center loss application. We performed experiments to validate the performance of our proposal using a hate speech detection dataset of manually labeled Wikipedia comments. Our experiments show that this method can outperform previous architectures for hate speech detection with an accuracy of 94.08% and an F1-score of 68.46% on the hate speech detection dataset.