基于大语言模型的财务报告指标抽取智能体方法

Agent-Based Financial Report Indicator Extraction with Large Language Models

摘要: 金融年报中的信息抽取因其复杂的PDF格式和超长上下文而极具挑战。传统的检索增强生成(RAG)方法受限于单步、静态的检索范式，一旦初始查询与文档表述不匹配，便容易失败。为解决这一检索脆弱性问题，本文提出了一个名为LedgerLens的多智能体协作框架。该框架借鉴人类分析师的认知模式，其核心研究智能体(Researcher Agent)通过“检索–分析–精炼”的迭代循环，在初步检索结果不佳时能够自主重构查询并进行多轮尝试，直至精准定位目标信息。在自建的银行年报问答数据集(BAR-QA)上的实验结果表明，LedgerLens在指标抽取任务中取得了94.1%的F1分数，并在大多数任务上实现了领先表现。研究结果证明，引入基于智能体的迭代查询优化机制，是突破传统RAG在复杂真实场景中检索瓶颈的有效途径。

Abstract: Information extraction from financial annual reports is highly challenging due to their complex PDF formatting and extremely long contexts. Traditional retrieval-augmented generation (RAG) methods rely on a single-step, static retrieval paradigm, which is prone to failure when the initial query does not align with document expressions. To address this retrieval vulnerability, this study proposes LedgerLens, a multi-agent collaborative framework. Inspired by the cognitive process of human analysts, the core Researcher Agent operates through an iterative “retrieve-nalyze-refine” loop, enabling it to autonomously reformulate queries and conduct multiple retrieval attempts until the target information is accurately located. Experiments on a self-constructed Bank Annual Report Question Answering dataset (BAR-QA) demonstrate that LedgerLens achieves an F1 score of 94.1% in the indicator extraction task and outperforms baselines on most tasks. These results indicate that introducing an agent-based iterative query optimization mechanism provides an effective solution to overcoming the retrieval bottlenecks of traditional RAG in complex real-world scenarios.