文章背景与核心概要
在构建生产级知识图谱时,将提取出的候选实体和关系写入图谱是一个高风险的过程,其中身份判定在本质上具有破坏性。错误的数据类型通常很容易在事后修正,但一旦两个不同的记录在属性合并后被归为一个单一身份,这种合并便是不可逆的,且不会留下任何错误痕迹。本文详细介绍了一个数据摄入与本体标记流水线,该流水线成功处理了 98,795 份政府文件,构建出一个包含 537,157 个实体 和 2,198,567 条关系 的强大生产级知识图谱。
该系统通过引入“记录-身份阶梯”(Record-Identity Ladder)来系统性评估实体同一性,摒弃了单纯依赖名称相似度的做法;同时采用“保守的自动化策略”,彻底杜绝了未经审查的自动合并,避免生产事故。此外,通过在多类标记中引入锚定证据,系统有效避免了虚假分类的产生,将虚假角色指派从 36 例大幅降至 4 例。论文还量化了图谱的一致性债务,并详细阐述了包含 48,403 个待处理提案 的人机协同维护队列。
Curate Before You Connect: Identity and Ontology Tagging in a Production Knowledge Graph
arXiv: 2608.10644 [cs.AI]
Submitted on: August 11, 2026
Authors: Vaibhav Dangaich, Kevin Lewis, Kundeshwar Pundalik
arXiv: 2608.10644 [cs.AI]
Submitted on: August 11, 2026
Authors: Vaibhav Dangaich, Kevin Lewis, Kundeshwar Pundalik
Summary
将提取出的候选实体和关系写入知识图谱是一个高风险的过程,其中身份判定在本质上具有破坏性。虽然错误的数据类型可以很轻松地在事后进行修正,但一旦两个不同的记录在属性合并后被归为一个单一身份,这种合并便是不可逆的——并且不会留下任何错误痕迹。
Writing extracted candidate entities and relationships into a knowledge graph is a high-stakes process where identity decisions are fundamentally destructive. While an incorrect data type can easily be corrected later, merging two distinct records under a single identity is irreversible once their properties have combined—leaving behind no error trail.
本文概述了一个摄入与本体标记流水线,该流水线旨在处理 98,795 份政府文件,并构建出一个稳健的生产级知识图谱,包含 537,157 个实体 和 2,198,567 条关系。该系统的核心亮点包括:
This paper outlines the ingestion and ontology-tagging pipeline designed to process 98,795 government documents into a robust production knowledge graph comprising 537,157 entities and 2,198,567 relationships. Key highlights of the system include:
- 记录-身份阶梯(The Record-Identity Ladder): 实体同一性不再依赖简单的名称相似度,而是通过结构化的标识符列、名称列、显示名称以及类型作用域位置构成的层级进行系统性评估。
- 保守的自动化政策(Conservative Automation Policies): 在经历了一次惨痛的生产事故(表面形式的名称合并污染了有效记录并删除了不相关的实体)之后,作者实施了严格的自动化阈值:实体消歧工具仅标记候选匹配项,完全避免了未经审查的自动化合并。
- 多类标记中的锚定证据(Anchored Evidence in Multi-Class Tagging): 作者发现了一种意想不到的证据不对称性——将实体名称视为实例标签而不是任意类型断言,可以防止系统凭空捏造错误的分类。要求提供具体的、有锚定的证据,使 enriched 样本中的错误角色指派从 36 减少到了 4,且所有这 4 个都得到了成功验证。
- 管理与 conformance 指标(Curation & Conformance Metrics): 本文量化了图谱的一致性债务,阐述了次要分类如何弥补父级分类错误的类,并详细介绍了不断增长的人机协同策划队列,其中包含针对 775 个已解决的人工决策 的 48,403 个待处理提案。
- The Record-Identity Ladder: Rather than relying on simple name similarity, entity sameness is systematically evaluated using a structured hierarchy of identifier columns, name columns, display names, and type-scoped positions.
- Conservative Automation Policies: Following a costly production incident where surface-form name merging corrupted valid records and deleted unrelated entities, the authors implemented a strict automation threshold: entity resolution tools only ever flag candidate matches, completely avoiding unvetted automated merges.
- Anchored Evidence in Multi-Class Tagging: The authors identified an unexpected evidence asymmetry—treating an entity name as an instance label rather than an arbitrary type assertion prevents the system from inventing false classifications. Requiring concrete, anchored evidence drastically reduced false role assignments from 36 down to 4 on an enriched sample, all of which were successfully verified.
- Curation & Conformance Metrics: The paper quantifies the graph’s conformance debt, illustrates how secondary classifications compensate for mis-parented primary classes, and details a growing human-in-the-loop curation queue containing 48,403 pending proposals against 775 resolved human decisions.
Article Details & Metadata
Article Details & Metadata
- 主要主题: 人工智能 (
cs.AI) - 引用方式:
arXiv:2608.10644 [cs.AI] - DOI: 10.48550/arXiv.2608.10644
- 许可协议: 知识共享 署名-非商业性使用-相同方式共享 4.0

- Primary Subject: Artificial Intelligence (
cs.AI)- Cite As:
arXiv:2608.10644 [cs.AI]- DOI: 10.48550/arXiv.2608.10644
- License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0
访问与全文链接
Access & Full-Text Links