Attentional Multi-Channel Convolution With Bidirectional LSTM Cell Toward Hate Speech Prediction

Abstract

Online social networks(OSNs) facilitate their users in real-time communication but also open the door for several challenging problems like hate speech and fake news. This study discusses hate speech on OSNs and presents an automatic method to identify hate messages. We introduce an attentional multi-channel convolutional-BiLSTM network for the classification of hateful content. Our model uses existing word representation techniques in a multi-channel environment having several filters with different kernel sizes to capture semantics relations at various windows. The encoded representation from multiple channels passes through an attention-aware stacked 2-layer BiLSTM network. The output from stacked 2-layer BiLSTM is weighted by an attention layer and further concatenated and passes via a dense layer. Finally, an output layer employing a sigmoid function classifies the text. We investigate the efficacy of the presented model on three Twitter-related benchmark datasets considering four evaluation metrics. In comparative evaluation, our model beats the five state-of-the-art and the same number of baseline models. The ablation study shows that the exclusion of channels and attention mechanism has the highest impact on the performance of the presented model.The empirical analysis analyzing the impact of different word representation techniques, optimization algorithms, activation functions, and batch size on the presented model ascertains the use of their optimal values.