Aspect-based Multilabel Classification of E-commerce Reviews Using Fine-tuned IndoBERT

In recent years, e-commerce has experienced rapid growth. A significant change in consumer behavior is marked by the ease of access and time flexibility offered by e-commerce platforms, as well as the existence of the review feature to assess products and services. However, with the ever-increasing number of reviews, consumers and store owners face challenges in sorting out relevant information. This research focuses on the multilabel classification of Indonesian e-commerce reviews. This research was undertaken because the application of multilabel classification, especially for e-commerce reviews in Indonesia, has received little attention. This research compares three classification models: IndoBERT, IndoBERT-CNN, and IndoBERT-LSTM, to determine the most effective model for multilabel aspect classification of customer reviews. The multilabel classification method was applied to determine the aspect categories of the reviews, such as product, customer service, and delivery, using different thresholds for evaluation. Results show that 0.6 threshold is optimal, with the IndoBERT-LSTM model as the best-performing model for the multilabel aspect classification of these e-commerce reviews. Optimal classification of the model enables more precise information extraction from customer reviews. This can be useful for e-commerce businesses to gain insight from the reviews they get from customers. This insight can be used to find out which aspects need to be improved from the e-commerce business which leads to increased customer satisfaction and trust.