跳转至

文章背景与核心概要

实体消歧（Entity Disambiguation, ED）是构建和利用知识图谱的核心环节，传统上通常被视为一个将候选检索与基于上下文的选择相结合的整体任务。本文探讨了解耦这两个子问题的显著优势。通过将无需训练的稀疏检索器（BM25）与强大的大语言模型（LLM）选择器配对，作者证明了训练专用检索器所带来的收益正在逐渐递减。

研究所提出的模块化流水线在 ZELDA 基准测试中取得了 SOTA（State-of-the-Art）的领先成果，在显著提升性能的同时，还能够通过选择性弃权（selective abstention）机制稳健地处理检索失败的情况。该论文已被第25届国际语义Web会议（ISWC 2026）研究赛道接受。

Select, Don't Train: The Benefits of Modular Entity Disambiguation with LLM-Based Selection

Executive Summary

Entity Disambiguation (ED) is a crucial process for building and utilizing knowledge graphs, traditionally treated as a monolithic task combining candidate retrieval and context-based selection. This paper explores the advantages of decoupling these two subproblems. By pairing a training-free sparse retriever (BM25) with a powerful Large Language Model (LLM) selector, the authors demonstrate that training specialized retrievers offers diminishing returns. The proposed modular pipeline achieves state-of-the-art results on the ZELDA benchmark, significantly improving performance while enabling robust handling of retrieval failures through selective abstention.

Publication Details

  • arXiv ID: arXiv:2608.27470 [cs.CL]
  • Authors: Fina Polat, Daniel Daza, Pengyu Zhang, Klim Zaporojets, Paul Groth
  • Submission Date: August 17, 2026
  • Conference Acceptance: Accepted at the 25th International Semantic Web Conference (ISWC 2026), Research Track
  • Primary Subject: Computation and Language (cs.CL)
  • Additional Subjects: Artificial Intelligence (cs.AI), Databases (cs.DB)

Abstract

Entity Disambiguation (ED) is a key task for constructing and using knowledge graphs. State-of-the-art neural approaches commonly model ED as a single task, although it consists of two distinct subproblems: retrieving candidate entities and selecting the correct one given context. Dual-encoder models optimize for both within a shared embedding space, forcing representations to balance high-recall retrieval with fine-grained selection, and they require trained retrievers, which are costly to maintain as knowledge graphs change.

While recent work has begun to combine retrievers with LLM-based selectors, the interplay between the two stages has not been studied systematically. In this paper, we present a systematic comparison of retrieval strategies for candidate generation under a shared LLM-based selection stage, combining sparse retrieval (BM25), Web KB search, and a state-of-the-art trained dense retriever with several open- and closed-source LLMs.

We show that, once selection is delegated to a capable LLM, training the retriever provides only modest additional value: * A fully training-free BM25 retriever paired with an LLM selector reaches a new state of the art on the ZELDA benchmark, raising inKB micro-F1 from 82.3 to 86.3 (+4). * Pairing the same LLM with a trained dense retriever reaches 88.5.

Decoupling retrieval from selection also exposes a limitation of current ED systems: when the correct entity is missing from retrieved candidates, they are forced to predict an incorrect entity. In contrast, our framework allows for abstention when retrieval failure is detected. In an evaluation setting that rewards correct abstentions, the training-free BM25 + LLM pipeline reaches 90.7 F1.

Full-Text & Resource Links

摘要

实体消歧（ED）是构建和使用知识图谱的关键任务。最先进的神经方法通常将 ED 建模为单一任务，尽管它包含两个不同的子问题：检索候选实体以及根据上下文选择正确的实体。双编码器模型在共享的嵌入空间中对这两者进行优化，迫使表征在保证高召回率的检索与细粒度选择之间进行平衡，并且它们需要经过训练的检索器——随着知识图谱的变更，维护这些检索器的成本很高。

尽管近期的研究已经开始将检索器与基于大模型的选择器结合起来，但这两个阶段之间的相互作用尚未得到系统性的研究。在本文中，我们在统一的基于大模型的选择阶段下，对用于候选生成的检索策略进行了系统性比较，结合了稀疏检索（BM25）、Web KB 搜索以及最先进的训练密集检索器，并测试了多个开源和闭源大语言模型。

我们表明，一旦将选择任务委托给功能强大的一流大模型，训练检索器所能带来的附加价值微乎其微： * 完全无需训练的 BM25 检索器与大模型选择器相结合，在 ZELDA 基准测试中创下了全新的 SOTA 记录，将 inKB micro-F1 从 82.3 提升至 86.3 (+4)。 * 将同一款大模型与经过训练的密集检索器配对，则可达到 88.5。

将检索与选择解耦，同时也暴露了当前 ED 系统的一个局限性：当检索到的候选对象中缺少正确的实体时，它们被迫预测一个错误的实体。相比之下，我们的框架允许在检测到检索失败时进行弃权。在一个对正确弃权予以奖励的评估设置中，无需训练的 BM25 + LLM 流水线达到了 90.7 的 F1 值。

Entity Disambiguation (ED) is a key task for constructing and using knowledge graphs. State-of-the-art neural approaches commonly model ED as a single task, although it consists of two distinct subproblems: retrieving candidate entities and selecting the correct one given context. Dual-encoder models optimize for both within a shared embedding space, forcing representations to balance high-recall retrieval with fine-grained selection, and they require trained retrievers, which are costly to maintain as knowledge graphs change.

While recent work has begun to combine retrievers with LLM-based selectors, the interplay between the two stages has not been studied systematically. In this paper, we present a systematic comparison of retrieval strategies for candidate generation under a shared LLM-based selection stage, combining sparse retrieval (BM25), Web KB search, and a state-of-the-art trained dense retriever with several open- and closed-source LLMs.

We show that, once selection is delegated to a capable LLM, training the retriever provides only modest additional value: * A fully training-free BM25 retriever paired with an LLM selector reaches a new state of the art on the ZELDA benchmark, raising inKB micro-F1 from 82.3 to 86.3 (+4). * Pairing the same LLM with a trained dense retriever reaches 88.5.

Decoupling retrieval from selection also exposes a limitation of current ED systems: when the correct entity is missing from retrieved candidates, they are forced to predict an incorrect entity. In contrast, our framework allows for abstention when retrieval failure is detected. In an evaluation setting that rewards correct abstentions, the training-free BM25 + LLM pipeline reaches 90.7 F1.

全文与资源链接

Full-Text & Resource Links