Orthogonal least square based feature selection for an automatic hate speech detection and classification

Abstract

Hate speech in social media is a growing issue nowadays that negatively affects the society and individuals within. Moreover, hate speech detection is a challenging task, due to the vast number of user data generated on a daily basis. This makes it difficult to review each comment made by a user. The major objective of this research is to effectively identify or detect hate speech, and classify the same, using Orthogonal Least Squares (OLS)-based feature selection. The different feature extraction approaches known as Term-Frequency-Inverse Data Frequency (TF-IDF), count vectorizer, global vector (GloVe) model, and aspect features, are used to extract frequent occurrences of the keywords, coverage of keywords and contextual meaning of words. The features chosen by the OLS are classified using Stacked Bidirectional Long Short-Term Memory with Multiple Attention Mechanism (SBiLSTM-MAM) to compute the attention among the hidden states and utterance embeddings. Thus, the developed OLS discards the unwanted features and concentrated on most significant features, that improves classification using SBiLSTM-MAM. Here, the OLS ranks the features according to its orthogonal significance, confirming that the chosen features are less collinear, thus minimizing the noise. From the result analysis, it clearly shows that the precisions gained by the proposed OLS-SBiLSTM-MAM, in the OLID and SOLID datasets, are 80.22% and 84.7% respectively, which are higher when compared to that of existing Softplus BiLSTM.