Perbandingan Long Short-Term Memory dan Bidirectional Long Short-Term Memory pada Analisis Sentimen Ulasan Wisata

Abstract

Tourism reviews on Google Maps can be used to understand visitors’ perceptions of destination quality, including aspects that often become sources of satisfaction or complaints. This study analyzes reviews of Pantai Alam Indah because the destination has a large number of reviews, diverse text forms, and an imbalanced sentiment distribution. The purpose of this study is to compare the performance of Long Short-Term Memory and Bidirectional Long Short-Term Memory in tourism review sentiment analysis. The data were collected through Google Maps scraping and processed through case folding, cleaning, tokenizing, stopword removal, and stemming. The final dataset consisted of 2,702 reviews, including 2,024 positive reviews and 678 negative reviews. To reduce the effect of class imbalance, the training process applied class weighting by assigning a higher weight to the negative class. Both models used a 128-dimensional embedding layer, a 64-neuron dense layer, a sigmoid activation function in the output layer, Adam optimizer, batch size of 32, 10 epochs, and a validation split of 0.2. In addition to sentiment classification, the reviews were grouped into cleanliness, facilities, price, and general aspects using a keyword-based rule-based approach. The evaluation results show that Bidirectional Long Short-Term Memory achieved an accuracy of 75.79%, precision of 76.33%, recall of 75.79%, and F1-score of 76.04%, while Long Short-Term Memory achieved an accuracy of 75.05%, precision of 75.48%, recall of 75.05%, and F1-score of 75.25%. The performance difference between the two models was relatively small, so the results should be interpreted carefully. The aspect analysis shows that the general aspect dominated positive sentiment, while cleanliness had the highest number of negative sentiments.