跳转至

文章背景与核心概要

大语言模型（LLMs）虽然能够生成流利的多文档摘要，但其传统的归因方式往往较为粗粒度（通常引用整个文档或大段文本），且多为事后生成（post hoc），导致单个陈述难以验证。为了解决这一痛点，本文引入了 CAMS（Claim-Anchored Multi-document Summarization，声明锚定多文档摘要）这一全新框架，它将归因视为生成的结构化属性，而非下游预测任务。

CAMS 的核心创新在于：首先将每个源文档分解为原子声明，并通过确定性方式将溯源从原文引用精准解析到确切的 Token 跨度；其次对多篇文档中的等效声明进行聚类，同时显式标记源间冲突；最后通过支持感知重写，确保生成的摘要句子能够附带清晰映射回源文本跨度的声明标识符。该研究在理论上明确区分了“溯源”（独立于模型准确率、对每个输出句子均成立的不变结构属性）与“faithfulness/忠实度”（通过选择和验证鼓励实现的目标）——这两个概念在端到端和事后系统中经常被混淆。

在评测方面，作者在 MultiNews、DiverseSumm 以及 WCEP 的零样本（zero-shot）设置下对 CAMS 进行了评估。结果表明，CAMS 不仅在摘要质量上匹配了强大的端到端和跨度归因基线，显著提升了忠实度和引用精确度，还将多源归因准确率从 38% 提升至 64%（且未增加引用源数量），并将每个声明的人工验证时间缩短了 3.4 倍。

Attributable by Construction: Claim-Anchored Provenance for Multi-Document Summarization

Large Language Models (LLMs) are capable of producing fluent multi-document summaries, but their attributions are traditionally coarse (referencing whole documents or large passages) and generated post hoc, making individual statements difficult to verify.

This paper introduces CAMS (Claim-Anchored Multi-document Summarization), a novel framework that treats attribution as a structural property of generation rather than a downstream prediction task.

Key Innovations of CAMS:

  1. Deterministic Provenance Resolution: Decomposes every source document into atomic claims, resolving their provenance deterministically from verbatim quotes down to exact token spans.
  2. Conflict Flagging & Claim Clustering: Clusters equivalent claims across multiple documents while explicitly flagging inter-source conflicts.
  3. Support-Aware Rewriting: Selects a salient, well-supported subset of claims and rewrites them so that every summary sentence concludes with claim identifiers mapping cleanly back to source spans.
  4. Theoretical Separation: Establishes a clear distinction between provenance (an invariant structural property holding for every emitted sentence independently of model accuracy) and faithfulness (an objective encouraged by selection and verification)—two concepts frequently conflated by end-to-end and post-hoc systems.

评估与结果

The authors evaluate CAMS on MultiNews, DiverseSumm, and in a zero-shot setting on WCEP. Using a strict two-regime evaluation protocol (separating reference-free citation quality from gold-aligned localization) audited by an independent support model:

  • 在摘要质量方面，与强大的端到端及跨度归因基线相媲美。
  • 显著提升了忠实度与引用精确度。
  • 将多源归因准确率从 38% 提升至 64%，且未膨胀引用源的数量。
  • 将每个声明的人工验证时间缩短了 \(3.4\times\)。
  • Matches strong end-to-end and span-attribution baselines regarding summary quality.
  • Improves faithfulness and citation precision significantly.
  • Raises multi-source attribution accuracy from 38% to 64% without inflating the number of cited sources.
  • Cuts human verification time per claim by \(3.4\times\).

资源与获取

  • 全文链接： 查看 PDF | HTML（实验性） | TeX 源码
  • 数据集与代码： 作者开源了源代码，并随 MultiNews 发布了约 32,0000 个“声明-引用-跨度”标注，作为可重复使用的细粒度归因资源。
  • Full-Text Links: View PDF | HTML (Experimental) | TeX Source
  • Dataset & Code: The authors release source code alongside roughly 320,000 claim--quote--span annotations over MultiNews as a reusable, fine-grained attribution resource.