A comparative study of deep learning algorithms for hate speech detection on Twitter

Abstract

Hate speech is an undesirable phenomenon with severe psychological and physical consequences. The emergence of mobile computing and Web 2.0 technologies has increasingly facilitated the spread of hate speech. The speed, accessibility and anonymity afforded by these tools present challenges in enforcing measures that minimise the spread of hate speech. The continued dissemination of hate speech online has triggered the development of various machine learning techniques for its automated detection. However, current approaches are inadequate because of further challenges such as the use of domain-specific language and language subtleties. Recent studies on automated hate speech detection have focused on the use of deep learning as a possible solution to these challenges. Although some studies have explored deep learning methods for hate speech detection, there are no studies that critically compare and evaluate their performance. This work investigates the use of deep learning algorithms as possible solutions to hate speech detection on Twitter. Three taxonomic classes of deep learning algorithms, namely, Traditional deep learning algorithms, Traditional algorithms with partial attention mechanism and Transformer models, which are entirely based on the attention mechanism, are evaluated for performance, using two publicly available corpora. One of the datasets contained 24 786 tweets annotated into three different classes, while the other dataset contained 2300 tweets annotated into two different classes. All tweets from the two datasets were first preprocessed to rid of them of characters and words deemed irrelevant to the classification decision, for instance, hashtags, stop words and punctuation marks. The preprocessed text was then transformed into feature vectors which were used as input for deep learning algorithms explored in this study. A series of experiments were performed to measure the performance of the deep learning algorithms in hate speech detection. The algorithms were tested on a wide spectrum of tweets containing different forms of hate speech. The efficacy of the deep learning algorithms was objectively evaluated using six state-of-the-art statistical evaluation metrics: precision, Fmeasure, recall, accuracy, Mathews correlation coefficient and area under the curve. The results from this study indicate that variations in parameters do not impact the efficacy of deep learning algorithms by the same proportions. The findings of this empirical study, therefore, provide deep-learning practitioners with a better understanding of the adaptation of robust deep-learning techniques for automated hate speech detection tasks.