Pengembangan Ekstensi Peramban Berbasis Fine-Tuning Indobert untuk Identifikasi Komentar Bermuatan Judi di Platform Youtube

Abstract

The rapid growth of social media has created opportunities for the spread of harmful content, including online gambling promotions in YouTube comment sections. Such promotional comments are often disguised using non-standard language, abbreviations, symbols, and obfuscated expressions, making them difficult to detect using conventional moderation approaches. This study aims to develop a browser extension based on a fine-tuned IndoBERT model for automatically identifying online gambling promotional comments on the YouTube platform in real-time. The dataset used in this study consists of secondary data collected from two public Indonesian-language YouTube comment datasets obtained from Kaggle. After data cleaning and deduplication, the final dataset comprised 18,658 comments categorized into two classes: gambling and non-gambling. The text preprocessing stage included Unicode normalization, case folding, removal of special characters, whitespace normalization, and post-preprocessing deduplication to prevent information leakage. The dataset was split into training, validation, and testing subsets using stratified sampling, while class weighting was applied to address class imbalance. The IndoBERT model (indobenchmark/indobert-base-p1) was fine-tuned using a learning rate of 3 × 10⁻⁵ and a batch size of 16. Two training scenarios were evaluated: Early Stopping with a patience value of 2 and Full Training up to 8 epochs. The final model was selected based on the highest validation F1-score and was obtained at epoch 7 in the Full Training scenario. The selected model was then deployed through a Flask-based API integrated with a Google Chrome browser extension. Evaluation on the test dataset showed that the model achieved an accuracy of 99.21%, precision of 99.20%, recall of 97.90%, F1-score of 98.54%, and ROC-AUC of 0.99. Error analysis indicated that most misclassifications were caused by obfuscated promotional patterns, abbreviated expressions, and ambiguous contextual comments. Functional evaluation through black-box testing confirmed that all browser extension features operated as expected. In addition, latency testing in a local environment showed an average response time of 200.286 ms. These results demonstrate that the fine-tuned IndoBERT model can be effectively implemented in a browser extension for real-time identification of online gambling promotional comments, providing a practical moderation support tool for Indonesian-language digital content.