<![CDATA[The rapid growth of user-generated content provides valuable insights into tourists’ perceptions of destinations. This study analyzes sentiment in Google Maps reviews of temple tourism destinations in Central Java using IndoBERT embeddings and a Bidirectional Long Short-Term Memory (BiLSTM) model. A total of 10,714 Indonesian-language reviews were collected through web scraping and processed through preprocessing, pseudo-labeling, embedding generation, and model training. To prevent data leakage, the dataset was divided into stratified training and testing sets, while Random OverSampling (ROS) was applied only to the training data. Since manually annotated labels were unavailable, sentiment categories were generated automatically using a pre-trained IndoBERT classifier. The BiLSTM model achieved 80.25% accuracy on the imbalanced dataset and approximately 95% accuracy against IndoBERT-generated pseudo-labels under balanced training conditions. Improvements in Macro F1-score and balanced accuracy indicate better recognition of minority classes. However, the results should be interpreted cautiously because pseudo-labeling and oversampling may affect performance. Overall, this exploratory study demonstrates the potential of IndoBERT and BiLSTM for Indonesian tourism sentiment analysis while highlighting the need for human-annotated data and stronger validation in future research.]]>
Copyrights © 2026
