Automatic Detection of Extremist Affiliations Using Deep Learning-Based Sentiment Analysis on Social Media Posts for Afaan Oromo

Abstract

In recent years, the Internet has become a global platform for communicating and disseminating information. Today, popular social media sites have a huge global reach and audience with many daily active users who interact and communicate through those social networks far apart from each other physically. Because there is no limitation that guides the users of the platform such as Facebook, twitter and YouTube, all users freely express their feelings, attitudes, beliefs and opinions towards various issues such as politics, social, religious, ethnic, business and etc. As a result, large volume of opinionated texts is produced on a regular basis; the comments might be positive or negative. It’s impractical and tedious to manually classify each and every review with human power into different class such as extreme, anti-extreme and neutral. Classifying the sentiments is used for the individuals, company as well as the government to make decision and action on the posts and comments to be removed that contains radical contents that may cause offence and conflict among the people. In the proposed work we designed and implemented deep learning-based sentiment analysis model for Afaan Oromo review. Dataset from three domains which is politics, religious, and ethnic is collected for the model. Totally 2410 reviews which is labeled with three classes extreme, anti-extreme and neutral labeled with 0,1, and 2 respectively prepared for the mode. Preprocessing techniques such as stop-word removal, tokenization, lower case conversion and other preprocessing techniques are used for cleaning the dataset. In the proposed model sentence level sentiment analysis is used. We implemented convolutional neural network (CNN) and long short memory (LSTM) with word2vec that classify the reviews into extreme, anti-extreme and neutral. The proposed model convolutional neural network achieve accuracy of 80% and long short-term memory achieved performance of 70%. According to our experiment convolutional neural network is good in classifying the review to the target class than the long short-term memory.