Integrated aspect extraction and sentiment classification for aspect-based sentiment analysis using fine-tuned indoBERT on indonesian e-commerce reviews

Abstract

The rapid growth of Indonesian e-commerce has generated vast volumes of consumer reviews, yet extracting actionable aspect-level sentiment from informal Indonesian-language texts remains challenging due to the limited availability of domain-specific Aspect-Based Sentiment Analysis (ABSA) models. This study aimed to develop and evaluate an integrated IndoBERT-based ABSA model that combines aspect extraction and aspect sentiment classification within a single framework, applied to Indonesian beauty product reviews. A corpus of 500 beauty product reviews was processed through aspect extraction, yielding approximately 10,000 aspect-level data points labeled as positive or negative. The IndoBERT model was fine-tuned with optimized hyperparameters. The model achieved 86% accuracy, 85.71% F1-score, and 88% balanced accuracy. Aspect-level evaluation revealed F1-scores of 100% for seller, 98% for product, and 86% for shipping. Inference throughput of 33,173 samples per second confirmed real-world deployment feasibility. These results demonstrate the effectiveness of integrated IndoBERT fine-tuning for ABSA on Indonesian e-commerce reviews and provide a foundation for enhancing data-driven marketing strategies in the beauty product sector.