Decoding User Trust in Crypto Wallets with a BERT–XGBoost Hybrid Model for Multilingual Phantom Review Analysis

The rapid expansion of decentralized financial applications has increased the importance of understanding user trust in crypto wallet platforms. This study examines trust expressions in multilingual Phantom Wallet reviews using a hybrid classification framework that integrates BERT-based contextual embeddings with an XGBoost model. A total of 12,422 English and Indonesian reviews were collected and processed to construct a multilingual dataset for trust analysis. Exploratory findings reveal a highly polarized distribution of user ratings, indicating that trust in crypto wallets is strongly influenced by clear satisfaction or dissatisfaction rather than moderate evaluations. Cross-linguistic analysis indicates that Indonesian users express a higher proportion of low-trust reviews compared to English users, suggesting greater sensitivity to transaction errors and perceived asset safety concerns. Lexical patterns demonstrate that positive trust is associated with usability and performance stability, while negative trust is primarily driven by system failures, delays, and missing balance incidents. The results confirm that the BERT–XGBoost hybrid model is well-suited for decoding trust-related signals by combining contextual semantic understanding with structured metadata. This study contributes to the broader discourse on digital trust within Web3 environments by demonstrating an effective multilingual machine learning approach for analysing user perceptions in decentralized financial technologies.