Aspect-Based Sentiment Analysis of Indonesian Electric Vehicles on Media Social

Abstract

The increasing adoption of electric vehicles (EVs) in Indonesia has generated diverse and unstructured public opinions across social media platforms. This study applies Aspect-Based Sentiment Analysis (ABSA) to classify Indonesian EV opinions across six predefined aspects: battery, design, price, performance, infrastructure, and general perception. The ABSA process was conducted using manual multi-label aspect-sentiment annotation rather than fully automatic aspect extraction, allowing one review to contain multiple EV aspects with corresponding sentiment labels. A total of 3,000 reviews were collected from YouTube, Instagram, and TikTok and used to evaluate three deep learning architectures: Bi-LSTM, IndoBERT, and Hybrid IndoBERT-BiLSTM. In the Hybrid architecture, IndoBERT was used as a Transformer-based contextual feature extractor, and the resulting contextual embeddings were passed into a Bi-LSTM layer to capture bidirectional sequential dependencies before final sentiment classification. The dataset was divided into training, validation, and testing sets using an 80:10:10 split. Model performance was measured using accuracy, precision, recall, and F1-score. The results show that IndoBERT achieved the highest average accuracy of 80.50%, followed by Hybrid IndoBERT-BiLSTM with 78.83% and Bi-LSTM with 71.67%. Although IndoBERT performed best overall, the Hybrid model showed competitive performance and better validation-loss stability in selected aspect-level contexts. These findings indicate the effectiveness of Transformer-based models for Indonesian EV sentiment analysis, while hybrid sequential modeling can provide a stable alternative for handling informal social media text.