Classification of User Reviews for Software Maintenance in Indonesian Language Using IndoBERT-BiLSTM (Case Study: MyPertamina)
Andreyan Rizky Baskara, Muhammad Ardhy Satrio Jati, Mutia Nadra Maulida, Yuslena Sari, Nurul Fathanah Mustamin, Eka Setya Wijaya · 2023
Useful app reviews play an important role for developers in maintaining app quality and performance. However, with the increase in application usage and the number of reviews provided, not all reviews are constructive, some reviews only contain insults or praise without providing constructive suggestions or feedback. Manually sifting through thousands of reviews takes a lot of time and effort. Therefore, there is a need for efficient and effective methods to identify useful reviews. This study used the IndoBERT word embedding method and the BiLSTM classifier for classification of useful reviews. Experimental results show that the IndoBERTBi-LSTM model with a learning rate configuration of 2e-5, dropout probability of 0.2, and batch size 16 achieves the best results with an accuracy value of 95.49% and shows an increase in accuracy of 1.16% compared to the fine-tuned IndoBERT model as classification method.