Aspect Extraction on Indonesia Beauty Product Review Using Pre-Trained Language Model

Currently, the beauty industry is experiencing a growth of 20.6% from 2021 to 2022. With this growth, there are currently many forums to discuss and review these beauty products, one of which is Female Daily. These reviews play an important role in influencing consumer purchasing decisions. According to brightlocal data on consumer reviews, 82
^%
of respondents tend to delay purchases if they find negative reviews. Therefore, reviews have a considerable impact on the perceived quality of the product. One way to get an inference or summary of reviews is with aspect-based sentiment analysis. Therefore, this study performs aspect extraction on beauty products using trained language models, namely IndoBERT and mBERT. The process includes labeling data, preprocessing, split data, data filtering, modeling, aspect prediction, and evaluation. The highest F1-score achieved by mBERT was 61%, while IndoBERT was 47%. In this study, an error analysis was conducted and found that some words in the reviews contained English which caused the mBERT F1-score to be higher than IndoBERT. The English words in the reviews include ‘cushion’, ‘matte’, ‘foundation’, ‘dewy’, and others. In addition, in the performance analysis for each aspect, the highest result was obtained by the price and texture aspect, which was higher at 65%, compared to the packaging aspect.