A Study on Latent Dirichlet Allocation Generative Model to Cryptocurrency Fraud Judicial Documents

In recent years, the incidence of cryptocurrency-related fraud has remained persistently high in Taiwan. To better understand the role of cryptocurrencies in fraudulent activities and to identify their criminal characteristics, this study applies text mining techniques to analyze court rulings from the past three years involving cryptocurrency-based fraud. Using the TF-IDF algorithm, keywords were extracted from judicial decisions to identify prominent patterns. The results indicate that cryptocurrency fraud primarily comprises three dimensions: transaction mechanisms, fraud syndicate operations, and the use of deceptive rhetoric and technological tools. Furthermore, the topic modeling via Latent Dirichlet Allocation (LDA) across multiple years reveals that while the overarching legal frameworks and fraudulent narratives have remained relatively consistent, the instruments used in fake investment schemes have evolved—from conventional tools such as bank accounts and passbooks to more complex financial products like debt instruments and bonds. The fraud schemes themselves have also shifted, with early cases in 2022 often involving impersonation of law enforcement officers, whereas later cases increasingly involve the recruitment of juveniles into fraud syndicates. Notably, in 2024, the tracing of financial flows emerged as a dominant theme, reflecting an increased allocation of resources by law enforcement toward blockchain-based transaction tracing. The findings of this study offer valuable insights for law enforcement agencies and policymakers, serving as a reference for developing crime prevention strategies and early warning systems.