Sentiment Analysis of TikTok Comments Using Fine-Tuned IndoBERT for Indonesian Brand Perception

Abstract

The growth of user generated content on social media has increased the need for automatic sentiment analysis of unstructured text. This research fine tuned an IndoBERT based architecture for ternary sentiment classification of 3,857 TikTok comments on Point Coffee, collected between January 2020 and February 2025. The procedure involved preprocessing, two stage labeling combining an Indonesian sentiment lexicon with annotator adjudication, stratified partitioning, class balancing through undersampling, and supervised fine tuning with AdamW for three epochs. On a held out partition of 1,070 comments the model yielded an accuracy of 94.49%, with a 95% Wilson interval of 92.95 to 95.70%, and macro averaged precision, recall, and F1 of 94.62%, 94.45%, and 94.41%. The error distribution is more informative than the totals: 55 of 59 misclassifications occur at the neutral boundary and only four invert polarity.