FinConRank: Financial Constraint-Aware Visual Reranking for Multimodal Document Retrieval

Abstract

Financial reports contain repeated layouts and dense tables, so visually plausible pages may report the wrong entity, metric, year, or accounting scope. We present FinConRank, a financial constraint-aware visual reranker adapted with page-level hard negatives mined from actual retrieval neighborhoods. A bilingual parser extracts query conditions, and conservative evidence rules distinguish observable conflicts from missing information. Pairwise low-rank adaptation (LoRA) produces a financial relevance score combined with ColQwen2, MonoQwen2-VL, and Metric-Aware evidence. We evaluate English and Chinese FinRAGBench-V retrieval using report-group nested query splits over a fixed document corpus; training negatives may expose pages from reports associated with held-out queries. Relative to the same-protocol MonoQwen2-VL baseline, four-score fusion improves normalized discounted cumulative gain at rank 10 (nDCG@10) and mean reciprocal rank at rank 10 (MRR@10) by 0.032 and 0.040 in English, and by 0.021 and 0.023 in Chinese. All eight original primary comparisons remain significant after report-clustered testing and Holm’s correction. A separately selected three-score fusion without FinConRank narrows the attribution claim: adding the financial score improves English ranking, whereas its incremental Chinese effect is inconclusive after correction. A matched hard-negative analysis and a stratified single-annotator audit provide complementary training-data evidence. These findings support improved top-ranked financial evidence retrieval within the evaluated fixed-corpus benchmark.