Detection of hate speech in twitter space using LSTM neural network

Abstract

With the exponential increase in internet usage in last decade, the way of communication has become more digitalized. This has driven to numerous positive outcomes . At the same time, it has brought many negatives too. The volume of destructive substance online, such as hate discourse, is huge. This intrigued within the scholastic community to investigate an automated tool for discovering the hate speech. It become very necessary to find out an automated technique to put a check on these hateful contents. In this project work I have proposed a deep learning model that helps to find out the hateful contents automatically . In this work , I tried out my experiment over four different languages . English [35] dataset contains 15,777 tweets classified over three different classes such as Non-Hate , Racism and Sexism .German[38] dataset has 3031 tweets in german language classified over two different classes Non – Hate and Hate . Italian[37] dataset has 3000 tweets in Italian classified over two different classes Non – Hate and Hate . Bengali[36] dataset has 3419 tweets in Bengali and classified over five different classes Geopolitical , political , personal , religious and gender abusive . My proposed method is based on RNN based LSTM deep neural network along with the FASTTEXT word embedding model . The best result for English dataset is obtained by FASTTEXT+LSTM method with an accuracy of 0.81825 . Also we have get a better result for Bengali dataset by FASTTEXT+LSTM method with an accuracy of 0.61988 comparing it to WITHOUT FASTTEXT+LSTM method . But for the Italian and German dataset we get the best result using WITHOUT FASTTEXT+LSTM method . In case of Italian dataset we get the accuracy of 0.78 and in case of German dataset we get the accuracy of 0.70.