DistilBERT-Based E-Commerce Sentiment Analysis

Abstract

The rapid advancement of digital technology has driven significant growth in
Indonesia’s e-commerce sector, with Shopee emerging as one of the largest
platforms generating millions of product reviews daily. These reviews contain
valuable consumer opinions that can be analyzed to assess customer
satisfaction, yet their massive volume makes manual analysis inefficient and
subjective. This study aims to develop an automated sentiment analysis model
using DistilBERT to classify Shopee product reviews into positive and
negative sentiments. The dataset comprises approximately 1 million English
language reviews covering various product categories, including electronics,
fashion, beauty, and household items. The research methodology involves text
preprocessing, tokenization using DistilBertTokenizerFast, and fine-tuning of
the DistilBERT model under multiple data-split ratios (90:10, 80:20, 70:30,
60:40). Experimental results demonstrate that DistilBERT achieved the
highest accuracy of 94.8%, outperforming baseline models such as Naïve
Bayes (88.4%) and SVM (89.6%). These findings confirm that DistilBERT
effectively maintains a balance between accuracy, precision, and recall while
offering high computational efficiency. This research contributes both
methodologically and practically by establishing DistilBERT as a
scientifically robust and resource-efficient solution for large-scale sentiment
analysis in Indonesia’s e-commerce environment.