FinCards: Structured Reranking for Financial Document QA
FinCards, a novel framework, enhances the selection of evidence in financial question answering related to lengthy corporate filings. It approaches the task as a constraint satisfaction problem by utilizing aligned schema fields for entities, metrics, fiscal periods, and numeric spans. The framework incorporates multi-stage tournament reranking alongside stability-aware aggregation, resulting in auditable decision traces. When evaluated on two benchmarks for corporate filing QA, FinCards significantly outperforms both lexical and LLM-based reranking baselines in early-rank retrieval.
Key facts
- FinCards is a structured reranking framework for financial document QA.
- It reframes evidence selection as constraint satisfaction under a finance-aware schema.
- The schema includes fields for entities, metrics, periods, and numeric spans.
- It uses multi-stage tournament reranking with stability-aware aggregation.
- FinCards produces auditable decision traces.
- It was evaluated on two corporate filing QA benchmarks.
- It improves early-rank retrieval over lexical and LLM-based reranking baselines.
Entities
—