<![CDATA[PT Kereta Api Indonesia (Persero), or PT KAI, is a key public transportation provider in Indonesia whose service quality significantly influences public perception. The proliferation of digital platforms allows users to voice their opinions on PT KAI’s services, rendering sentiment analysis a vital tool for evaluating public satisfaction. This study analyzes public sentiment toward PT KAI by benchmarking the performance of Long Short-Term Memory (LSTM) and IndoBERT models. We collected 82,044 user reviews from various digital platforms including X (formerly Twitter), Instagram, TikTok, YouTube, Facebook, and Google Play Store specifically focusing on the “Access by KAI” application. Data labeling utilized the InSet Lexicon, and the dataset was partitioned into training and testing sets at an 80:20 ratio. During the initial training phase, the LSTM model exhibited overfitting, evidenced by a significant disparity between training and validation performance. To mitigate this, several optimization strategies were implemented, including architecture simplification, dropout, L2 regularization, and early stopping. Post optimization, the LSTM model achieved 95.98% accuracy, while IndoBERT reached 96.15%. Although the 0.17% performance gap is marginal, IndoBERT demonstrated superior stability in processing informal social media registers without extensive normalization. Practically, this suggests that IndoBERT is more robust for real-time monitoring systems where data variety is high, whereas LSTM remains a viable, lightweight alternative for resource-constrained environments. This research contributes to computational linguistics by providing a robust performance comparison of transformer-based vs. recurrent architectures in processing the informal registers of the Indonesian language.]]>
