Detecting Islamophobic Hate Speech On Social Media Using Semi-Supervised Graph Convolutional Networks

Abstract

Islamophobia hate speech is expressed with a generalized negative attitude and behavior toward Muslims and Islam. Speaking out against Muslims has a detrimental effect on how people view Islam. Hate speech towards Muslims has significantly increased on social media during the past few years. This nature of rhetoric encourages violence and prejudice against the Muslim community as well as some violent Muslim reactions. This paper introduces Text Graph Convolutional Networks (TextGCN) model for semi-supervised text classification to identify Islamophobic content on social media. Leveraging a dataset of 1,617 annotated tweets, we applied a semi-supervised approach to classify a larger corpus of 28,000 tweets. This model captures intricate relationships between words, documents, and their interconnections within the text through a heterogeneous graph structure. By using two convolutional layers on this graph, TextGCN achieves state-of-the-art performance, surpassing existing methods with a test accuracy of 93.58% using BERT embeddings and 91.38% using Word2Vec+Doc2Vec embeddings.