Detection of Extreme Political Sentiments in Pakistan on Social Media

Abstract

The proliferation of political extremism on social media has adverse effects on not only the individuals who are targeted but also on society at large. It causes great damage to the hosting platform as well where such content is being shared. Even though notable research work has been done on sentiment analysis and classification in both academia and industry, an effective and robust tool to detect and classify political extremism on various social media platforms is still a challenge. Previous research work had largely focused on detecting general hate speech on social media via binary classification. But, considering the diverse nature of extremism, binary classification does not suffice the purpose. In this research, we have studied existing solutions and after finding their limitations, we have developed a multi-class and multi-lingual model that detects and distinguishes between neutral, moderate, and strong political extremist content. For training our model, we collected a data set of around 10,000 tweets from prominent political parties and politicians in Pakistan. We used the latest pre-trained BERT model and machine learning classifiers like Support Vector Machine, Random Forest, Naıve-Bayes, and Stochastic Gradient Descent to analyze and detect different classes of extremism. The highest accuracy we achieved is 89% in binary classification and 86% in multi-class classification using the Term Frequency-Inverse Document Frequency word embedding and SVM classifier. It is hoped that the results of this thesis will provide researchers and organizations with a viable solution to detect and classify extreme political sentiments.