Detection of Cyberbullying in Social Media to Control Users' Mental Health Issues Using Recurrent Neural Network Architectures

Abstract

Cyberbullying is a kind of bullying that takes place via social media with digital gadgets. It may cause significant, long-lasting trauma and difficulties with mental and physical issues. This condition may also result in difficulties in personal lives. Because of this, detecting and reporting such offensive posts may help avoid the harmful repercussions of cyberbullying. Some studies are available in this context, and it was found that deep learning algorithms are more efficient in detecting abusive texts on social media. Some studies were carried out on deep learning algorithms. The issue was not solved, and more studies are needed to find the most efficient model to control this problem. It was observed from the literature survey that the performance of Recurrent Neural Network (RNN) architectures is good. Therefore, the objective of this study was to find and develop an efficient model to detect cyberbullying texts on social media. For this research, the Twitter dataset was used, which was uploaded by Munki Albright in Kaggle. The dataset consists suspicious, cyberbullying, hate and suicidal classifications. This research was focused on cyberbullying texts. Sentiment analysis was implemented to classify the cyberbullying posts. For that, sexism (Class 2), racism (Class 1) and either (Class 0) classification were analyzed. Researchers applied the LSTM, GRU and Bidirectional LSTM architectures of RNN to perform the sentiment analysis. All models gave somewhat similar results. When we consider f1, GRU shows a better F1 score for all the classes when compared to others (Class 0 was 95%, Class 1 – 70%, and Class 0 was 56%). GRU shows the best results for this data set compared to other models. Based on the Accuracy, all three architectures were obtained at around 90%. However, GRU has outperformed the other two models in accuracy as well. After that, researchers developed an ensemble model using all three models. In the ensemble, each model has been given weight. Since the GRU is the best performing model, it was given 0.4, and the other two models were assigned 0.3 each. The ensemble model performed well in F1 measurement (57%) and accuracy (91.17%). When we compared the previous works against the proposed model, our ensemble model obtained the highest values and performed well in detecting cyberbullying in sentiment analysis.