<![CDATA[Bot-generated comments in online news discussions can reduce discussion quality and complicate content moderation, particularly when automated comments resemble natural human language. This study aims to evaluate IndoBERT for classifying Indonesian online-news comments into bot and human classes and to compare its performance with Random Forest as a conventional machine-learning baseline. The dataset consists of 1,000 comments collected through web scraping from six Indonesian news and discussion platforms, namely Detik.com, Kompas.com, Kompasiana.com, Kaskus.co.id, Quora.com, and CNN Indonesia. The dataset contains 500 bot comments and 500 human comments. Comments were labeled using rule-guided manual annotation based on repetitive patterns, external promotional links, textual similarity, templated structures, contextual irrelevance, and posting-frequency patterns. The data were preprocessed and divided into 70% training data and 30% evaluation data. IndoBERT was fine-tuned for binary classification using three training epochs. The evaluation produced 95.67% accuracy, 91.56% precision, 100% recall, 95.59% F1-score, and 0.9164 MCC, while Random Forest achieved an F1-score of 91.84%. The results indicate that IndoBERT achieved higher F1-score than Random Forest on the dataset and evaluation procedure used in this study. However, the findings should be interpreted within the study setting because external, temporal, cross-platform validation, repeated runs, and broader baseline comparisons were not performed. The contribution of the study lies in the empirical evaluation of IndoBERT for Indonesian online-news bot-comment classification, while the web application serves as an implementation artifact.]]>
Copyrights © 2026
