This paper presents the first case study applying Large Language Models to the German Central Bank's process of verifying securities eligibility for collateral, shifting from traditional Named Entity Recognition to a generative Information Extraction pipeline. The approach decomposes the task into extraction, normalization, and interpretation to handle noisy text and bilingual content more effectively.
- The system achieves high precision of up to 91% in document-level eligibility verification.
- It utilizes an LLM-as-a-judge evaluation methodology for semantic assessment rather than location-based metrics.
- The model exhibits a conservative operating profile that minimizes false acceptance of ineligible assets.
This method reduces the resource intensity of manually verifying lengthy, semi-structured prospectuses by overcoming limitations associated with OCR noise and linguistic variance.