Automatic hate speech detection using aspect based feature extraction and Bi-LSTM model

The various social media platforms are used for easy access of information from the distinctive field that might constitute an offensive discussion. Therefore, existing studies are examined to reduce offensive harassment cases online. The spread of hate speech is growing with the ubiquity and anonymity through the means of social media for many years. Thus, the increase in demand showed an automated model for the detection of hate speech. The existing models utilized deep learning models failed to analyze the syntax and grammar or even modify original data’s meaning due to its complex patterns. Therefore, the present research work utilizes an Activation Function known as Soft-plus in Bidirectional Long Short Term Memory (Bi-LSTM) models for hate speech detection which helps the network to learn complex patterns in the data. The proposed Soft-plus Bi-LSTM learned the c complex patterns present in the network data and the activation function made an end decision that should be fired out into the next neuron. The classification results showed that the proposed Soft-plus Bi-LSTM classified the reviews as abusive or non-abusive speech. The results obtained better precision values of 60.09% when compared to the existing models Auto-Encoder and Multi-task learning model showed a precision of 53.9% and 55.7% respectively.