Sentiment Classification of TikTok Reviews on Almaz Fried Chicken Using IndoBERT and Random Oversampling

Abstract

The socio-political context surrounding the Indonesian Ulema Council's Fatwa No. 83 of 2023, which catalyzed a significant consumer shift, necessitates an accurate measure of public sentiment toward alternative local brands like Almaz Fried Chicken. Analyzing real-time consumer discourse on the challenging TikTok platform, the study utilized a final dataset of 4,374 unique comments to overcome the inherent problem of dataset imbalance and linguistic informality. The core method involved a seven-stage quantitative approach: data collection, preprocessing, sentiment labeling, data splitting (70:15:15), Random Oversampling (ROS), IndoBERT fine-tuning, and evaluation. This pipeline fine-tuned IndoBERT, a Transformer-based model, integrated with ROS applied exclusively to the training data. Evaluation demonstrated that ROS significantly reduced model bias and enhanced performance: Overall Accuracy increased by 2.0% (from 91% to 93%), and the Macro F1-Score improved by 3.4% (from 0.87 to 0.90). Most critically, the F1-Score for the minority Negative sentiment class surged from 0.78 to 0.84, confirming ROS's effectiveness in accurately detecting critical feedback. These findings provide timely, data-driven insights into brand perception amidst the boycott campaign and establish a robust, reliable IndoBERT-ROS methodology for advanced sentiment monitoring in dynamic social media environments.