用于混合长文档的多粒度证据检索方法研究

Research on Multi-Granularity Evidence Retrieval Methods for Hybrid Long Documents

摘要: 混合长文档(Hybrid Long Documents, HLDs)在问答任务中存在效率低下和证据分散的挑战，现有方法直接将整个文档作为大语言模型(Large Language Models, LLMs)的输入会引发噪声干扰和事实幻觉现象。为解决这些问题，本文提出以相关证据句子为核心的知识支持策略，并设计了三阶段处理框架进行证据句子检索，第一阶段采用分层检索策略，通过双阶段过滤机制根据用户声明从文档中精准定位相关证据段落；第二阶段针对包含表格的段落设计序列化处理方法；第三阶段通过段落去噪筛选证据句子，与用户声明结合起来辅助大语言模型生成答案。此外，为了推进领域研究，我们构建了首个中文可行性研究报告数据集，在数据集上进行的实验表明：我们模型超越了传统的基线模型，其准确率达到80.4%。

Abstract: Hybrid long documents face challenges of low efficiency and scattered evidence in question-answering tasks. Existing methods directly use the entire document as input to a large language model, which can cause noise interference and factual hallucination phenomena. To address these issues, this paper proposes a knowledge support strategy with relevant evidence sentences as the core, and designs a three-stage processing framework for evidence sentence retrieval. In the first stage, a hierarchical retrieval strategy is adopted to accurately locate relevant evidence paragraphs from documents based on user statements through a two stage filtering mechanism. In the second stage, a serialization processing method is designed for paragraphs containing tables. In the third stage, evidence sentences are filtered through paragraph denoising, and combined with user statements to assist the large language model in generating answers. In addition, to advance research in the field, we constructed the first Chinese feasibility study report dataset. Experiments on the dataset show that our model surpasses the traditional baseline model with an accuracy of 80.4%.