Klasifikasi Jenis Ulasan Konsumen Produk Kecantikan Lokal Menggunakan Pendekatan BERT (Bidirectional Encoder Representations from Transformers)

Abstract

Product reviews can be information for potential consumers to find out about the product to be purchased. Reviews can be useful as a reference and consideration for consumers before buying a product. There are various methods for providing product reviews, namely through writing, social media, photos and videos or a combination of various media. A widely used method is to provide reviews through writing. Someone can give their opinion and can be analyzed to find out the type of review submitted including neutral sentences, complaints suggestions or praise. Thus, this research was conducted to classify the types of reviews contained in the text. In this study, the BERT (Bidirectional Encoder Representations from Transformers) approach method was used. 2000 data in the form of beauty product reviews collected through the SOCO Review site were used in the study. Then the dataset is annotated into 4 classes of review types, namely neutral, complaints, suggestions, and praise before entering the classification process. The pre-processing stages in this study include case folding, data cleaning, tokenization, stopword removal, and normalization. Model building uses the BERTBASE architecture, namely IndoBERT-base-p1. The hyperparameters used in this research are batch size 32, learning rate 3e-6, and 5 epochs. The results achieved in this study resulted in a fairly good accuracy of 90%.