A Unified Framework for Knowledge-Intensive Numerical Reasoning over Financial Document

Numerical reasoning in financial document analysis constitutes a fundamental challenge for corporate financial report understanding, drawing increasing attention in both academia and industry. However, current approaches suffer from semantic misalignment in multi-hierarchical tables, reasoning disruptions from insufficient integration of financial metric formulas, and sensitivity of the model’s reasoning results to the order of evidence. To address these challenges, we propose a unified framework for knowledge-intensive numerical reasoning over financial documents. Within this framework, we introduce a Triplets-based Multi-Stage Tabular-Textual Hybrid Evidence Retrieval (THER) method to resolve semantic misalignment by converting multi-hierarchical tables into triplet representations. Furthermore, we propose the Fine Grained Knowledge Injected Chain-of-Thought (FGKI-CoT) method to enhance numerical reasoning by explicitly integrating financial conceptual formulas into the reasoning path. Building on FGKI-CoT, we introduce the Evidence Order Sampling based Self-Consistency (EOSC) method, which mitigates the model’s sensitivity to evidence order by altering the input evidence sequence. Experiments demonstrate that our framework enables a 1.5B-parameter language model to outperform GPT-3.5-turbo by 3.95% in numerical reasoning on the Multihiertt Dev dataset. Additionally, we conduct supplementary experiments to further explore the impact of table representations and reasoning step expressions on the numerical reasoning performance of language models.