OPTIMALISASI MODEL TRANSFORMER UNTUK IDENTIFIKASI CYBERBULLYING DI MEDIA SOSIAL PADA KONDISI KETERBATASAN SUMBER DAYA KOMPUTASI

Luky Vianika Sari, NIM.: 23206052010 (2025) OPTIMALISASI MODEL TRANSFORMER UNTUK IDENTIFIKASI CYBERBULLYING DI MEDIA SOSIAL PADA KONDISI KETERBATASAN SUMBER DAYA KOMPUTASI. Masters thesis, UIN SUNAN KALIJAGA YOGYAKARTA.

Preview	Text (OPTIMALISASI MODEL TRANSFORMER UNTUK IDENTIFIKASI CYBERBULLYING DI MEDIA SOSIAL PADA KONDISI KETERBATASAN SUMBER DAYA KOMPUTASI) 23206052010_BAB-I_IV-atau-V_DAFTAR-PUSTAKA.pdf - Accepted Version Download (7MB) | Preview
	Text (OPTIMALISASI MODEL TRANSFORMER UNTUK IDENTIFIKASI CYBERBULLYING DI MEDIA SOSIAL PADA KONDISI KETERBATASAN SUMBER DAYA KOMPUTASI) 23206052010_BAB-II_sampai_SEBELUM-BAB-TERAKHIR.pdf - Published Version Restricted to Registered users only Download (7MB) | Request a copy

Abstract

Cases of cyberbullying on social media continue to rise and have serious impacts on mental health, particularly among teenagers. Although various artificial intelligence–based approaches have been developed to detect cyberbullying, many of them still rely on complex model architectures, complicated preprocessing steps, and high computational demands, making them less ideal for use in resource-constrained environments. This study aims to optimize a transformer-based model for multi-class cyberbullying classification on social media content under limited computational resources. In this research, the authors evaluate and compare the performance and training effort of various transformer-based models (BERT, RoBERTa, XLNet, HateBERT), hybrid architectures (RoBERTa+BiLSTM, RoBERTa+CNN, BERT+BiLSTM), and a majority-voting ensemble model. The dataset used contains more than 47,000 English-language tweets labeled into six categories: religion, gender, ethnicity, age, other cyberbullying, and non-cyberbullying. All models in this study were fine-tuned using default hyperparameter settings, without complex tuning, external feature engineering, or advanced ensemble techniques such as stacking. The majority-voting ensemble combining RoBERTa, BERT, and XLNet achieved the best performance, with an accuracy of 85.01%, a macro F1 score of 85.04%, and a weighted F1 score of 84.88%. Although it does not surpass the accuracy of highly specialized systems from previous studies, this approach emphasizes simplicity and operational efficiency. These results indicate that transformer-based ensembles can effectively classify various types of cyberbullying with a lightweight and practical configuration, making them suitable for deployment in resource-limited environments.

Share this knowledge with your friends :

Actions (login required)

View Item