Automatic Detection of Toxic South African Tweets Using Support Vector Machines with N-Gram Features

Toxic South African corpus is not available to detect toxic tweets such as offensive, hate, bullying and violent tweets. But there are some offensive and hate speech corpora, mostly in English, which have been used to detect toxic tweets. This paper focuses on automatic detection of toxic South African tweets using a reliable English corpus. The review of text classification models has shown that Support Vector Machines have very often outperformed other classic machine learning algorithms, while word and character n-gram features have performed well with varying prediction performances in different contexts. This paper therefore evaluated the performance of different parameter settings of Support Vector Machines and n-gram features for detection of toxic South African tweets, with a view to hybridize the best among the classifiers. Different combinations of word and character n- gram features were used for the classification. The results show that the Support Vector Machine classifier with set of unigram and bigram as well as set of character n-gram with length sizes from 3 to 7 perform best. By combining the classifiers, the accuracy and F-measure improve from the initial highest Accuracy and F-Measure scores of 0.9085 and 0.94, respectively to 0.9095 and 0.95. The comparison of our results with the performance of previous work on the English corpus shows that our model is reliable.