Fine-Tuning a Multilingual Translation Model for Financial Crime Data

This manuscript presents a focused study on fine-tuning a multilingual-to-English translation model for financial crime detection, addressing critical gaps in domain-specific machine translation (Johnson et al., 2017; Tiedemann & Thottingal, 2020). The model, based on Helsinki-NLP’s opus-mt-mul-en, was adapted using a curated multilingual dataset of 50,000 synthetic financial crime records, incorporating patterns identified in AML/CFT research (FATF, 2022). The fine-tuned model was evaluated across six languages—Hindi, French, Spanish, Chinese, Arabic, and Bengali—demonstrating significant improvements in BLEU scores (Lin, 2004), validating its efficacy for financial content translation. These advancements align with emerging applications of NLP in financial surveillance (Singhal et al., 2020) while mitigating technical debt risks in production systems (Sculley et al., 2015).