Digitalization of gold pawn services: Customer satisfaction and innovation challenges analyzed using Indonesian bidirectional encoder representations from transformers

Abstract

Background: The rapid digitalization of financial services has transformed customer interaction patterns, making user reviews a critical source for evaluating service quality. In digital gold pawn services, understanding customer sentiment is essential for sustainable digital transformation. While deep learning models often outperform conventional methods, their application in the Indonesian gold pawn sector remains limited. This study analyzes user sentiment toward the Tring! by Pegadaian application and evaluates a Transformer-based model's effectiveness in capturing customer satisfaction and complaints. Methods: This research employed a quantitative experimental approach using 21,413 Google Play Store reviews. Classification was performed using a pre-trained IndoBERT (indobert-base-p1) model, fine-tuned over three epochs with class weighting to address severe data imbalance. Performance was evaluated using accuracy, precision, recall, F1-score, and confusion matrix analysis. Findings: The results demonstrate that the model achieved an accuracy of 86.36% and a weighted F1-score of 0.8840. A recall of 79.46% for the negative sentiment class confirms the model’s capability to effectively detect customer complaints. Analysis revealed a dominance of positive sentiment related to ease of use and transaction speed, while negative sentiment highlighted issues regarding system stability and cost transparency. These findings show that Transformer architectures effectively capture semantic nuances in informal Indonesian text. Conclusion: IndoBERT-based sentiment analysis provides a reliable framework for evaluating digital gold pawn services. By achieving high accuracy and recall, the model effectively transforms unstructured feedback into actionable insights, supporting data-driven decision-making for service optimization in financial institutions. Novelty/Originality: This research applies a class-weighted IndoBERT model to a large-scale Indonesian gold pawn dataset, addressing severe class imbalance inherent in financial apps. It positions sentiment analysis as a real-time early warning system (EWS) for state-owned enterprises (BUMN) to monitor digital transformation challenges and maintain competitive advantages in the fintech landscape.