Enhancing Cyberbullying Detection in Bengali Language Through Hybrid Ensemble Method: The Code Against Hate Initiative
Md. Zunead Abedin Eidmum, Bakhtiar Muiz, Rakib Hossen, Anichur Rahman, Md Jahidul Islam · 2024
Celebrities and public figures, especially women, face a new challenge every day on social media comments known as cyberbullying, which raises the need for an effective cyberbullying detection system. Detecting cyberbullying on social media is crucial to preventing public harassment. While cyberbullying is an everyday event, it can result in depression and suicide. Timely identification can remove those comments early and help law enforcement take necessary actions in certain cases. However, the investigation process is very hard and time-consuming, especially for the Bengali language, as Bengali is a large and diversified language. Therefore, the main goal of a cyberbullying detection system in the Bengali language is to classify cyberbullying comments accurately. This paper introduces a state-of-the-art hybrid model that uses a unique feature extraction method that combines Bidirectional Encoder Representations from Transformers (BERT) embedding with transformed numerical and categorical features and one machine learning model named Support Vector Machines (SVM) and two deep learning models named multi-head self-attention based Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM) with Particle Swarm Optimization (PSO). To improve the performance further, we ensemble those models together. The proposed model with LSTM and Bagging method achieves a high accuracy rate of 99.26%. The proposed ensemble model outperforms the existing works and sets a new standard for Bengali Language cyberbullying detection methods in social media comments.