<![CDATA[Purpose – This study aimed to compare the performance of IndoBERT and Bidirectional Long Short-Term Memory (BiLSTM) in classifying public sentiment toward the Indonesian National Police (INP) based on YouTube comments. This study sought to identify a robust sentiment classification model to support text-based public perception monitoring, particularly under a highly imbalanced sentiment distribution.Method – YouTube comments were collected using the YouTube Data API. A total of 8,268 raw comments were obtained, and 7,197 comments were retained as the final dataset after preprocessing, automatic pseudo-labeling, and confidence filtering using a 0.5 threshold. To address concerns regarding threshold selection, an additional sensitivity analysis was conducted using confidence thresholds of 0.65 and 0.75. The experiment applied a dual-track preprocessing pipeline, cost-sensitive learning through class-weighted loss, bootstrap confidence interval analysis, and BiLSTM preprocessing ablation.Findings – The results show that IndoBERT achieved stronger performance than BiLSTM. IndoBERT obtained an accuracy of 92.92% and a Macro-F1 Score of 0.8548, whereas BiLSTM achieved an accuracy of 76.11% and a Macro-F1 Score of 0.6124. Bootstrap analysis showed a Macro-F1 difference of 0.2424, with a 95% confidence interval of 0.1870 to 0.2959, indicating that IndoBERT’s advantage was statistically significant. Sensitivity analysis also confirmed that IndoBERT consistently outperformed BiLSTM across all the tested thresholds.Research Implications – The findings indicate that IndoBERT is more suitable for Indonesian sentiment classification in public perception monitoring than other models. However, because the dataset labels were generated using a BERT-based classifier, the evaluation may contain architectural circularity that favors the IndoBERT model. Future studies should use human-annotated gold-standard data and broader cross-platform validations.Originality – This study provides a comparative evaluation of transformer-based and recurrent models using sensitivity analysis, bootstrap testing, cost-sensitive learning, and pre-processing ablation under imbalanced sentiment conditions.]]>
