Financial Voucher Analysis with LVMs and Financial LLMs

This work presents a novel computational framework for automated financial voucher analysis that combines large vision-language models (LVMs) and specialized financial large language models (LLMs) into a multi-stage, scalable pipeline. Leveraging transformer-based architectures, including a domain-specific 7 billion parameter financial language model, the system converts unstructured voucher data into structured representations to uncover meaningful financial patterns. The framework uses statistical methods like Z-score analysis and coefficient of variation to allow anomaly identification in three dimensions: amount-based, temporal, and pattern-based. Extraction and analysis processes are formalized through mathematical models and implemented algorithmically, with built-in fallback mechanisms to handle model uncertainty and ensure statistical validity. Our comparative evaluation demonstrates that specialized financial LLMs provide a 19.9% performance improvement over general-purpose LLMs, with particular gains in financial terminology recognition (+18.9%) and numerical accuracy (+13.2%). Evaluation on financial voucher data from different companies demonstrates the effectiveness of the framework in cash flow analysis, tax compliance, expense categorization, and anomaly detection, with a 15.7% reduction in false positives for anomaly identification. This methodology advances the state of automated financial document analysis, with significant implications for auditing, compliance, and business intelligence.