Transformer-based sentiment modeling for identifying cross country fintech perception gaps

Abstract

Conventional sentiment analysis lacks granularity for capturing detailed user experiences and cross-country comparative insights in digital finance. This study identifies and maps perception gaps among ShopeePay users in Indonesia and Thailand using a topic-informed sentiment analysis pipeline inspired by aspect-based sentiment analysis (ABSA) principles. Adopting the knowledge discovery in databases (KDD) framework, over 170,000 web scraped reviews were preprocessed, automatically labeled through pseudo labeling, and balanced using random oversampling. To mitigate pseudo-label reinforcement, manual validation on 500 reviews per country achieved agreement rates of 98.80% (Indonesia) and 99.20% (Thailand) with Cohen’s Kappa above 0.97. The fine-tuned DistilBERT model achieved accuracies of 97.65% (Indonesia) and 98.38% (Thailand), though these figures should be interpreted within the pseudo-labeled evaluation context. Significant perception gaps were revealed: Thai users showed lower satisfaction with transactions (67.6% negative) while Indonesian users were more positive (93.0% positive). Process time received negative dominance in both countries, with Indonesia at 78.2% and Thailand at 50.8% negative. These findings demonstrate that user satisfaction is shaped by local infrastructure and cultural contexts, providing strategic insights for regional fintech development.