Analisis Sentimen Ulasan Aplikasi Transportasi Online menggunakan IndoBERT dan Explainable AI (LIME)
TL;DR
It is demonstrated that explicit and implicit linguistic complexity is an important determinant of Transformer model performance and contributes to Indonesian-language sentiment analysis by integrating LIME to provide transparent explanations of model predictions for online transportation reviews.
Abstract
Gojek and Grab are two leading online transportation platforms in Indonesia, receiving millions of user reviews on the Google Play Store. Previous studies analyzing reviews of both applications have generally produced final sentiment classifications without transparently explaining why a review is categorized into a particular sentiment. This study integrates the IndoBERT model with the Explainable Artificial Intelligence (XAI) method, Local Interpretable Model-agnostic Explanations (LIME), to address this gap. The IndoBERT model (indobert-base-p1) was fine-tuned on a dataset of Gojek and Grab reviews collected using google-play-scraper, with semi-automatic labeling based on star ratings. Evaluation results show that the model achieved 90.57% accuracy and a weighted F1-Score of 90.60%, substantially exceeding IndoBERT performance in the implicit hate speech domain, which reached 55%. These findings demonstrate that explicit and implicit linguistic complexity is an important determinant of Transformer model performance. LIME analysis identified the word “memudahkan” as the most dominant feature, with a weight of +0.61, in driving positive sentiment classification. This study contributes to Indonesian-language sentiment analysis by integrating LIME to provide transparent explanations of model predictions for online transportation reviews.