A Hybrid Deep Learning Approach Leveraging Word2Vec Feature Expansion for Cyberbullying Detection in Indonesian Twitter

Abstract

Twitter, a social media platform that enables users to generate, post, update, and peruse brief messages known as tweets, unfortunately, is frequently misused for circulating negative content encompassing cyberbullying. The detrimental effects of cyberbullying on the mental well-being of victims are profound, with extreme cases culminating in suicide due to severe stress. Consequently, preventive measures, inclusive of the development of a cyberbullying detection system for Twitter, are imperative. This study introduces a hybrid deep learning approach, incorporating feature expansion with Word2Vec and feature extraction with TF-IDF, for constructing a cyberbullying detection system tailored to the Indonesian language on Twitter. A sequence of test scenarios was executed on a system developed using a dataset of 29,085 Indonesian tweets. The outcomes of this study demonstrate that the highest accuracy was achieved by the CNN-LSTM hybrid model with an accuracy of 79.26%, and the LSTM-CNN hybrid model with an accuracy of 79.48%. These findings substantiate that the amalgamation of hybrid models, Word2Vec for feature augmentation, and TF-IDF for feature extraction, yields superior accuracy compared to other deep learning models. Consequently, this study has succeeded in identifying cyberbullying on Twitter, contributing to the development of a healthier social media environment for users.