Aggression Detection in Twitter Data Using Transformer-Based Convolutional Neural Network Model

Abstract

Online environments facilitate the increase of anti-social behaviors in people's social interactions. Behaviors such as hate speech, cyberbullying, and trolling have increased significantly, especially in recent years, with the widespread use of social media. Detection of aggression and hateful speech is an important step in reducing and preventing cyberbullying. Cyberbullying is defined as comments made on social media to harm other individuals by using hateful, offensive, rude, humiliating, and sarcastic expressions. It is slow and expensive to try to achieve this with human control with the existence of rapidly growing data, so cyberbullying can be stopped by automatic detection of aggression. In this study, the automatic determination of aggression detection via Cyber-Trolls, which is the Twitter dataset, is discussed. A coder named LMTweets was trained on 20001 tweets to extract the features of the dataset. The extracted features are given as input to the convolutional neural network model to classify the text as aggressive / non-aggressive. In addition, three classification algorithms, namely Na & iuml;ve Bayes, Support Vector Machine, K-Nearest Neighbors, were applied. In addition, three transformer models, BERT, XLNet, and ULMFIT were applied along with the Convolutional Neural Network, Long Short-Term Memory, and Gated Recurrent Unit three learning algorithms. Python, Keras API, and Tensorflow are used together in the proposed model. The performance parameters obtained in the experimental results were determined as accuracy, precision, recall, F1-score, and AUC, and it was revealed that the LMTweets + CNN model performed better among all the models used.