A Deep Learning Model for Classifying the Hate and Offensive Language in Social Media Text

Abstract

Recently, we had introduced a model for identifying and removal of toxic content from twitter, using an Information Retrieval (IR) model SOIR (Semantic query Optimization-based Information Retrieval). Based on lexical and semantic analysis, SOIR identifies the class labels of tweets. The result demonstrates the superiority of the SOIR model. This model is accurate but social media is a big data problem and a significant amount of time and memory is required. In this paper the deep learning technique is used to process large-scale social media text data. First uses Natural Language Processing (NLP) based feature extraction to create four different sets of training samples i.e. TF-IDF-based features, POS Tagged Features, a reduced feature vector of POS and the combined vector of TF-IDF and POS tagged features. The deep Convolutional Neural Networks (CNN) is used to train the model and to classify hate and offensive language. The dataset has been obtained from Kaggle. The performance in terms of training accuracy, validation accuracy, training loss and validation loss has been measured with the time complexity. In addition, the class-wise Precision, Recall, F1-score, and Mean accuracy have also been investigated. From experimental results, we found TF-IDF and POS-based combined features provide superior performance.