Fine-Tuned IndoBERT for Aspect-Based Multi-Label and Multi-Class Sentiment Analysis on Indonesian Tourism Reviews
Abstract
Aspect-Based Sentiment Analysis detects each aspect in a text and labels its sentiment, making complex opinions easier to understand. Most Indonesian tourism sentiment research focuses on overall sentiment and simplifies aspect detection to a single-label task with only a limited number of aspect categories, which fails to capture the multiple aspects and diverse opinions often expressed in tourist reviews. This paper develops a fine-tuned IndoBERT model for multi-label aspect detection and multi-class sentiment classification on Indonesian tourism reviews. The proposed method comprises several stages, namely data collection, data labeling, data pre-processing, data transformation, model training, and evaluation. Our experiment shows that the proposed model achieves a micro-averaged F1-score of 78.99% for multi-label aspect category detection, with the highest F1-score obtained for Price and Value (84.55%) and the lowest for Crowds and Atmosphere (72.22%). For multi-class aspect-level sentiment classification, the model achieved an overall accuracy of 82.88% with the best performance on the Positive class (F1-score 90.12%) and the lowest performance on the Neutral class (F1-score 49.85%). Error analysis shows that ambiguous opinions, mixed sentiment, and complex user intent, particularly in reviews containing many aspects as challenges. Overall, the fine-tuned IndoBERT framework effectively extracts aspect-specific sentiment patterns from Indonesian tourism reviews.