<![CDATA[Online reviews on Google Maps are a strategic resource for restaurant management, but a single review often praises one aspect while criticizing another, so document-level sentiment cannot reveal which aspect drives satisfaction or complaints. This study applies Aspect-Based Sentiment Analysis (ABSA) to Soramen restaurant reviews through a two-stage span-level pipeline—aspect-term extraction (ATE) then sentiment classification (ASC)—across four aspects (FOOD, SERVICE, PRICE, PLACE) with IndoBERT, evaluating parameter-efficient Low-Rank Adaptation (LoRA) against full fine-tuning to map critical service aspects. The data comprise 1,742 original reviews (3,930 spans) expanded to 3,061 reviews through controlled synonym-replacement augmentation and split using StratifiedGroupKFold with five folds. The results show that LoRA r=16 attains a Joint Partial F1 of 0.7744 with only 0.48% of parameters trained, comparable to full fine-tuning (0.7727); the difference is not statistically significant (paired t-test, p=0.82), so LoRA r=16 performs on par with full fine-tuning while using roughly 200 times fewer parameters. Analysis of the labeled corpus—in which (with negative reviews intentionally oversampled for training and therefore do not represent population proportions—) shows FOOD as the most frequently mentioned aspect (64.4% of reviews), while SERVICE and PRICE carry the largest share of complaints (around 36% each), and FOOD and PLACE remain dominated by positive sentiment. This study confirms LoRA as an efficient alternative for Indonesian ABSA while producing managerial recommendations based on complaint themes for Soramen.]]>
