<![CDATA[Aspect-Based Sentiment Analysis (ABSA) on hotel reviews faces significant challenges regarding semantic complexity and severe class imbalance, particularly in low-resource languages like Indonesian. This study evaluates the effectiveness of fine-tuning IndoBERT, a pre-trained Transformer model, to address these issues by benchmarking it against classical statistical methods (TF-IDF) and static embeddings (Sentence-BERT). Utilizing the HoASA dataset, the experiment implements a Random Oversampling strategy at the text level to mitigate data sparsity in minority classes. Empirical results demonstrate that the fine-tuned IndoBERT significantly outperforms baselines on the majority of aspects, achieving a global accuracy of 97% and macro F1-score of 0.92. Granular per-aspect analysis reveals that the model’s self-attention mechanism captures linguistic context robustly in tangible aspects (e.g., wifi, service), yet faces persistent challenges in highly ambiguous aspects such as smell (bau) and general. Statistical significance tests (Paired t-test and Wilcoxon) confirm that the performance gains over baselines are statistically significant (p < 0.05) and not due to random chance. The study concludes that leveraging contextual representations from IndoBERT, combined with data balancing strategies, offers a superior and statistically robust solution for handling linguistic variations and class bias in the Indonesian hospitality domain.]]>
Copyrights © 2025
