Evaluating Machine Learning Techniques for Detecting Offensive and Hate Speech in South African Tweets

In recent times, South Africa has been witnessing insurgence of offensive and hate speech along racial and ethnic dispositions on Twitter. Popular among the South African languages used is English. Although, machine learning has been successfully used to detect offensive and hate speech in several English contexts, the distinctiveness of South African tweets and the similarities among offensive, hate and free speeches require domain-specific English corpus and techniques to detect the offensive and hate speech. Thus, we developed an English corpus from South African tweets and evaluated different machine learning techniques to detect offensive and hate speech. Character n-gram, word n-gram, negative sentiment, syntactic-based features and their hybrid were extracted and analyzed using hyper-parameter optimization, ensemble and multi-tier meta-learning models of support vector machine, logistic regression, random forest, gradient boosting algorithms. The results showed that optimized support vector machine with character n-gram performed best in detection of hate speech with true positive rate of 0.894, while optimized gradient boosting with word n-gram performed best in detection of hate speech with true positive rate of 0.867. However, their performances in detection of other threatening classes were poor. Multi-tier meta-learning models achieved the most consistent and balanced classification performance with true positive rates of 0.858 and 0.887 for hate speech and offensive speech, respectively as well as true positive rate of 0.646 for free speech and overall accuracy of 0.671. The error analysis showed that multi-tier meta-learning model could reduce the misclassification error rate of the optimized models by 34.26%.