Implementation of IndoBERT, IndoBERTa, and XLM-RoBERTa for Sentiment Classification of Alfagift and KlikIndomaret User Reviews

Sentiment analysis of mobile application reviews is a critical component of business decision-making in Indonesia's digital retail sector. This study implements and evaluates three Transformer-based pre-trained language models IndoBERT, IndoBERTa, and XLM-RoBERTa for three-class sentiment classification (positive, neutral, negative) on user reviews of the Alfagift and KlikIndomaret applications scraped from Google Play Store covering the period 2023–2026. The research dataset comprises 9,600 Alfagift reviews and 9,100 KlikIndomaret reviews collected via the Google Play Store API, subsequently manually annotated using a standardized annotation guideline. Each model was evaluated in both base and large variants with fine-tuning employing adaptive learning rate scheduling and dropout regularization. Experimental results demonstrate that XLM-RoBERTa-large achieves the highest performance, with accuracy of 94.12% and 93.56% on the Alfagift and KlikIndomaret datasets, respectively, surpassing IndoBERT-large (92.87%/91.43%) and IndoBERTa-large (91.94%/90.34%). In-depth analysis of confusion matrices, training curves, and per-class F1-score distributions reveals that multilingual XLM-RoBERTa excels at handling linguistic variation in informal Indonesian, code-mixing, and abbreviations commonly found in app user reviews. This research contributes to the comparative understanding of Indonesian-language and multilingual BERT-based models for the e-grocery domain and provides a benchmark baseline applicable to future research.