A Comparative Study of Embedding Techniques and Classifiers for Aspect-Based Sentiment Analysis of Shopee Reviews

Abstract

E-commerce platforms like Shopee generate massive volumes of user reviews that contain valuable insights about products, services, and user experiences. Aspect-Based Sentiment Analysis (ABSA) enables fine-grained sentiment classification by identifying sentiment polarity toward specific aspects such as product quality, pricing, delivery, and application performance. This study presents a comprehensive comparative analysis of different embedding techniques and classification models for ABSA on Indonesian Shopee reviews. We evaluate three embedding approaches: FastText, GloVe, and BERT embeddings, combined with four classification models: Support Vector Machine (SVM), Convolutional Neural Network (CNN), BERT, and IndoBERT. Our experiments focus on five key aspects: product, price, delivery, application, and general sentiment. The results demonstrate that FastText embeddings combined with IndoBERT classifier achieves the highest accuracy of 91.59%, while BERT embeddings show more balanced performance across different classifiers. The findings provide valuable insights for e-commerce platforms seeking to implement effective sentiment analysis systems for Indonesian market understanding. Keywords - Aspect-Based Sentiment Analysis, FastText, GloVe, BERT, IndoBERT