Deep Learning Hybrid Models for Multilingual Cyberbullying Detection: Insights from Bangla and Chittagonian Languages

Cyberbullying is a serious problem that affects online communication, and stopping it is essential to building safer digital spaces. The goal of this study is to address the problem of identifying cyberbullying content within the different linguistic contexts of Bangla and Chittagonian languages by utilizing deep learning techniques, with a particular focus on hybrid multilingual models. We provide a thorough analysis of both traditional machine learning (ML) models, which in our study reached accuracy values between 0.63 and 0.711 (with SVM as the highest performer), and deep learning (DL) models, which produced accuracy values between 0.69 and 0.811 (with CNN as the top performer) when trained and tested on both languages together. We also propose a set of combined hybrid network-based models, which improve accuracy. Among the models implemented, including BiLSTM+GRU (achieving an accuracy of 0.799), CNN+LSTM (with an accuracy of 0.801), CNN+BiLSTM (yielding an accuracy of 0.78), and CNN+GRU (attaining an accuracy of 0.804), especially the most complex model ((CNN+LSTM)+BiLSTM) achieved the highest score of 0.82. In comparison, mBERT, a popular multilingual transformer model, reached 0.79 accuracy. This research emphasizes high-performing hybrid neural networks’ significance in multilingual contexts, particularly for low-resource languages, contributing to effective cyberbullying detection in diverse populations.

Перевод пока недоступен