Detection of Hate Speech using BERT and Hate Speech Word Embedding with Deep Model

Abstract

ABSTRACT There is an increased demand for detecting online hate speech, especially with the recent changing policies of hate content and free-of-speech right of online social media platforms. Detecting hate speech will reduce its negative impact on social media users. A lot of effort in the Natural Language Processing (NLP) field aimed to detect hate speech in general or detect specific hate speech such as religion, race, gender, or sexual orientation. Hate communities tend to use abbreviations, intentional spelling mistakes, and coded words in their communication to evade detection, which adds more challenges to hate speech detection tasks. Word representation from its domain will play an increasingly pivotal role in detecting hate speech. This paper investigates the feasibility of leveraging domain-specific word embedding as features and a bidirectional LSTM-based deep model as a classifier to automatically detect hate speech. This approach guarantees that the word is assigned its negative meaning, which is a very helpful technique to detect coded words. Furthermore, we investigate the use of the transfer learning language model (BERT) on the hate speech problem as a binary classification task as it provides high-performance results for many NLP tasks. The experiments showed that domain-specific word embedding with the bidirectional LSTM-based deep model achieved a 93% f1-score, while BERT achieved 96% f1-score on a combined balanced dataset from available hate speech datasets. The results proved that the performance of pre-trained models is influenced by the size of the trained data. Although there is a huge variation in the corpus size, the first approach achieved a very close result compared to BERT, which is trained on a huge data corpus, this is because it is trained on data related to the same domain. The first approach was very helpful to detect coded words while the second approach achieved better performance because it is trained on much larger data. To conclude, it is very helpful to build large pre-trained models from rich domains specific content in current social media platforms.