Deteksi Cyberbullying Pada Teks Bilingual Menggunakan Bidirectional Long Short-Term Memory
TL;DR
The experimental results demonstrate that the BiLSTM model with the RMSProp optimizer is effective for detecting cyberbullying in bilingual Indonesian and English texts.
Abstract
The increasing use of social media not only provides various benefits but also contributes to the spread of cyberbullying. Detecting cyberbullying on social media is challenging because users frequently communicate in Indonesian, English, or a combination of both languages. In addition, previous studies have generally focused on detecting cyberbullying in a single language, limiting their ability to accommodate the characteristics of bilingual text. This limitation may lead to failures in detecting cyberbullying comments accurately and promptly, potentially causing psychological harm to victims. Therefore, an automated detection system capable of understanding the characteristics of bilingual text is needed. This study aims to develop a BiLSTM model for detecting cyberbullying in Indonesian and English texts. A bilingual dataset consisting of 21,308 Indonesian and English text samples was used to train the BiLSTM model. The experimental results show that the choice of optimizer affects model performance, with RMSProp outperforming Adam and SGD, achieving an accuracy of 96.01%, a precision of 96.03%, a recall of 96.01%, and an F1-score of 96.01%. These results demonstrate that the BiLSTM model with the RMSProp optimizer is effective for detecting cyberbullying in bilingual Indonesian and English texts.