Analisis Sentimen Ulasan Produk Kecantikan Menggunakan Bi-LSTM dan Weighted FastText Embedding

Abstract

The rapid expansion of the beauty product sector in Indonesia has contributed to a growing volume of consumer reviews, which hold significant potential for extracting insights into user perceptions toward products. This study seeks to conduct sentiment analysis on beauty product reviews by employing the Bidirectional Long Short-Term Memory (Bi-LSTM) algorithm, coupled with FastText word embeddings weighted through an Inverse Document Frequency (IDF) scheme. The data utilized in this research were sourced from the Female Daily platform. The research process encompasses text preprocessing, word representation using FastText, IDF-based word weighting, and Bi-LSTM model training with hyperparameter tuning involving the number of LSTM units, dropout rate, and learning rate. The evaluation results on the testing data reveal that the model achieved an accuracy of 85.96%. The optimal hyperparameter configuration consisted of 96 LSTM units, a dropout value of 0.3, and a learning rate of 0.0008. The best-performing model was subsequently deployed into a web-based system for automated sentiment analysis of beauty product reviews.