FinCARDS: Card-Based Analyst Reranking for Financial Document Question Answering
Researchers have introduced FinCARDS, a novel reranking framework designed to improve the accuracy and stability of question answering over long financial documents. Unlike existing methods that prioritize semantic relevance, FinCARDS reframes evidence selection as a constraint satisfaction problem using a finance-aware schema. This approach represents filing chunks and questions with aligned schema fields, enabling deterministic matching and auditable decision traces. Experiments on two benchmarks show FinCARDS significantly enhances early-rank retrieval and reduces ranking variance without requiring model fine-tuning. AI
IMPACT Enhances accuracy and auditability for financial document QA systems, potentially improving compliance and analysis.