Aspect-Based Sentiment Analysis (ABSA) of Ventela Shoe Reviews on TikTok Shop Using Fine-Tuned IndoBERT

Abstract

The massive volume of consumer reviews on the social commerce platform TikTok Shop makes it difficult for local shoe brands such as Ventela to understand consumer perception in a structured manner, while Indonesian-language Aspect-Based Sentiment Analysis (ABSA) studies on this platform remain very limited. This study aims to apply fine-tuned IndoBERT for aspect-based sentiment classification and to measure consumer perception of four product aspects, namely Comfort, Design, Durability, and Price. Using a computational experiment approach, 1,000 reviews were collected, automatically annotated using a lexicon-based method with negation handling, restructured into 706 review-aspect pairs and divided using an 80:20 stratified split, and used to train and compare three models: TF-IDF with Logistic Regression, TF-IDF with Linear SVM, and fine-tuned IndoBERT. Testing on 142 test samples shows that fine-tuned IndoBERT is superior, achieving an Accuracy of 0.8521 and an F1-Macro of 0.7813 and surpassing both baselines on four of five primary metrics. Analysis of 706 review-aspect pairs identifies Design (75.6% positive) and Price (71.8% positive) as the main strengths, while Comfort (32.7% negative) and Durability (30.8% negative) emerge as improvement areas related to sizing and the quality of adhesive and stitching. This study enriches Indonesian ABSA literature in the social commerce domain and delivers a ready-to-use web-based simulator built with Gradio to facilitate periodic consumer-perception monitoring for data-driven decision-making processes.