Multi-aspect Extraction in Indonesian Reviews Through Multi-label Classification Using Pre-trained BERT Models

Aspect extraction automatically identifies and categorizes specific aspects mentioned in the text to enable fine-grained opinion analysis in sentiment analysis. While previous studies have successfully extracted aspects, they often focused on a single-aspect per review, overlooking the presence of multiple aspects within sentences. This limitation has affected capturing a complete user opinion, thus posing a challenge despite the complexity of computations and annotations in supervised learning. In this study, we address the task of extracting multiple aspects from Indonesian reviews based on multi-label classification using Bidirectional Encoder Representations from Transformers (BERT) by implementing pre-trained models. BERT is chosen due to the ability to capture contextual information through bidirectional encoder mechanisms and the capacity to catch complex word and sentence relationships. In the experiment, we conducted the tests with various Indonesian pre-trained BERT models to enhance the performance of multi-aspect extraction on Indonesian hotel reviews. Our findings indicate that Indonesian-BERT-1.5G pre-trained model can improve the classifier performance and achieve an impressive F1-score of 0.84 with the lowest loos of 0.3029 for Indonesian reviews data.