跳转至

文章背景与核心概要

历史实体链接（Historical Entity Linking）旨在将历史文本中的实体提及映射到知识库中，而大型语言模型（LLM）在该领域展现出了巨大的应用潜力。然而，现有的偏好优化技术通常在每个训练样本中仅依赖一个单一的负面候选（single negative candidate），这种做法丢弃了其他检索到的负面候选所包含的宝贵信息。

为了克服这一局限性，本文作者引入了多负样本直接偏好优化（Multi-Negative Direct Preference Optimisation, MDPO）。这是一种基于参考的模型配对目标函数，它能够针对给定提及所检索到的所有有效拒绝候选来评估正确实体。实验表明，MDPO 在处理 NIL 提及（无对应实体）、语义歧义、光学字符识别（OCR）噪声以及复杂的历史实体名称时表现出显著提升，为大模型在历史文本处理中的应用提供了高效且实用的优化方案。

Beyond Single-Negative Preference: Multi-Negative DPO for LLM-Centric Historical Entity Linking

arXiv ID: arXiv:2609.07379 [cs.CL]
Submitted: 7 September 2026
Authors: Tien Nam Nguyen, Emanuela Boros, Ahmed Hamdi, Adam Jatowt, Mickaël Coustaty, Antoine Doucet

Beyond Single-Negative Preference: Multi-Negative DPO for LLM-Centric Historical Entity Linking

arXiv ID: arXiv:2609.07379 [cs.CL]
Submitted: 7 September 2026
Authors: Tien Nam Nguyen, Emanuela Boros, Ahmed Hamdi, Adam Jatowt, Mickaël Coustaty, Antoine Doucet

Summary

Large Language Models (LLMs) show great potential in historical entity linking, yet current preference optimization techniques typically rely on a single negative candidate per training instance. This approach discards valuable information present in other retrieved negative candidates.

To address this limitation, the authors introduce Multi-Negative Direct Preference Optimisation (MDPO)—a reference-based pairwise objective that evaluates the correct entity against all valid rejected candidates retrieved for a given mention.

Key Highlights:

  • Methodology: MDPO maintains the standard Bradley-Terry formulation of DPO while fully utilizing the complete set of candidate instances through masked, length-normalized sequence scores.
  • Evaluation: Tested across datasets like hipe-2020 and newseye, covering historical newspaper texts in French, German, English, Swedish, and Finnish.
  • Performance: MDPO outperforms traditional supervised fine-tuning and single-negative DPO, demonstrating notable improvements with NIL mentions, semantic ambiguity, optical character recognition (OCR) noise, and historically complex entity names.
  • Insights: Error analysis reveals that candidate retrieval remains the primary bottleneck for end-to-end historical entity linking, proving that multi-negative candidate incorporation offers a straightforward yet impactful boost to LLM performance.

Summary

Large Language Models (LLMs) show great potential in historical entity linking, yet current preference optimization techniques typically rely on a single negative candidate per training instance. This approach discards valuable information present in other retrieved negative candidates.

To address this limitation, the authors introduce Multi-Negative Direct Preference Optimisation (MDPO)—a reference-based pairwise objective that evaluates the correct entity against all valid rejected candidates retrieved for a given mention.

Key Highlights:

  • Methodology: MDPO maintains the standard Bradley-Terry formulation of DPO while fully utilizing the complete set of candidate instances through masked, length-normalized sequence scores.
  • Evaluation: Tested across datasets like hipe-2020 and newseye, covering historical newspaper texts in French, German, English, Swedish, and Finnish.
  • Performance: MDPO outperforms traditional supervised fine-tuning and single-negative DPO, demonstrating notable improvements with NIL mentions, semantic ambiguity, optical character recognition (OCR) noise, and historically complex entity names.
  • Insights: Error analysis reveals that candidate retrieval remains the primary bottleneck for end-to-end historical entity linking, proving that multi-negative candidate incorporation offers a straightforward yet impactful boost to LLM performance.

Paper Metadata

Paper Metadata

Citation & References

Citation & References

License Icon Reference:
license icon

License Icon Reference:
license icon