Detecting Hate Speech on Social Media Using Deep Learning Techniques

Abstract

Abstract Hate speech is a recurring issue on social media platforms identified as an attack against a specific group of people based on certain common characteristics. As online data is created at a very fast rate by users, it has now become a daunting task to manually moderate the comments of users containing hate speech in a bid to reduce its negative effects on a platform. Previous works have been able to create models capable of detecting hate speech with good accuracy on hate speech detection on user comments and posts (known as tweets) on Twitter social media platform. Despite the good results obtained, this kind of models perform poorly when exposed to tweets that contained clever wordings, alternate spellings and rare words. Therefore there is a need to improve the model for the detection of hate speech in user comments on social media in order to address these problems. An ensemble model was developed from two baseline classifiers, NBSVM (Naive Bayes Support Vector Machine ) and LSTM (Long Short Term Memory); combining the power of two well- known performing models from machine learning and deep learning using FastText embeddings from Facebook to improve hate speech detection even when clever wordings, alternate spellings and out of vocabulary words are used. This work was able to improve on the current state of the art hate speech detection by considering OOV (Out Of Vocabulary) words, clever and alternative spellings of words in developing a model that performed better than previous research works in detecting hate speech. The developed ensemble model proved to be able to detect hate speech even when clever wordings, alternate spellings and rare words were used in tweets. There was also an increase in the performance of the model’s hate recall (77%) as compared to the existing popular work of Davidson et. al, (2017) hate recall (61%).