Analisis Sentimen Berbasis Aspek pada Ulasan Aplikasi Redbus Indonesia: Fine-Tuning Indobert

Abstract

User reviews of online travel agent applications contain information about satisfaction and complaints regarding various services. However, general sentiment analysis cannot identify the specific service aspects underlying these evaluations. This study aims to identify aspects and sentiments in reviews of the redBus Indonesia application, evaluate the performance of IndoBERT after fine-tuning, and implement the classification results in a web-based dashboard. Reviews were collected from the Google Play Store for the period from May 6, 2023, to May 6, 2026, resulting in 4,463 reviews. Manual annotation was conducted by three anotators using a multi-label scheme covering Ticketing, Information Channels, and Travel Experience. After data selection, 1,977 reviews containing at least one aspect label were used for modeling. Annotation consistency was assessed using Fleiss’ Kappa, which produced values ranging from 0.6476 to 0.7462, indicating substantial agreement. Data processing included case folding, cleaning, normalization, back-translation, data splitting using MultilabelStratifiedShuffleSplit with an 80:10:10 ratio, and tokenization using the IndoBERT tokenizer with a maximum sequence length of 160 tokens. The indobenchmark/indobert-base-p2 model was then fine-tuning to classify six aspect– sentiment labels. The test results achieved a subset accuracy of 72.91%, macro precision of 81.09%, macro recall of 87.80%, and macro F1-score of 83.99%. The macro F1-score was close to the best validation score of 84.01%, indicating relatively consistent model performance. The Flask-based dashboard successfully presented classifications for manually entered reviews and Google Play Store reviews, while black-box testing showed that all functions operated as expected. These findings indicate that fine-tuning IndoBERT can effectively classify aspect-based sentiment in redBus Indonesia reviews and present the results through an informative dashboard.