Multilabel classification sentiment analysis on Indonesian mobile app reviews

Abstract

Mobile applications continue to evolve to satisfy the users. For that, the developers need to understand user feedback for improvements. Indonesia, one of the countries with the most mobile app users, has many textual mobile app reviews that may be processed and analyzed. Understanding the value of mobile app reviews requires understanding the value of sentiments and emotions to create more appropriate features to satisfy the users. To acquire a more accurate analysis of user reviews, it is important to detect sentiments that are closely associated with human emotion values due to the nature of multilabeled data. This research classifies the sentiments and emotions in Indonesian textual mobile app reviews, which are multilabel and multiclass in the form of 3 sentiments, namely positive, negative, and neutral, paired with 6 emotions, namely anger, sad, fear, happy, love, and neutral. We employ the Transformers architecture model, which includes two monolingual (a generic English and an Indonesian) and a multilingual pre-trained models with the results: bidirectional encoder representations from transformers (BERT) base uncased (micro avg. F1-score=0.69, precision=0.68, recall=0.70, receiver operating characteristic-area under the curve (ROC-AUC)=0.78), IndoBERT base uncased as best result (micro avg. F1-score=0.77, precision=0.78, recall=0.76, ROC-AUC=0.85), and multilingual BERT (M-BERT) base uncased (micro avg. F1-score=0.72, precision=0.73, recall=0.71, ROC-AUC=0.82).