3.1.3 实体与关系 → 知识图谱
5 / 8The use of an LLM to extract entities, relationships, and claims is a form of abstractive summarization – these are meaningful summaries of concepts that, in the case of relationships and claims, may not be explicitly stated in the text. The entity/relationship/claim extraction processes creates multiple instances of a single element because an element is typically detected and extracted multiple times across documents.
使用大语言模型抽取实体、关系和断言,是一种生成式摘要:这些内容是对概念的有意义概括;对于关系与断言而言,相应信息可能并未在原文中明确写出。由于同一个元素通常会在不同文档中被多次识别和抽取,实体、关系和断言的抽取过程会产生同一元素的多个实例。
In the final step of the knowledge graph extraction process, these instances of entities and relationships become individual nodes and edges in the graph. Entity descriptions are aggregated and summarized for each node and edge. Relationships are aggregated into graph edges, where the number of duplicates for a given relationship becomes edge weights. Claims are aggregated similarly.
在知识图谱抽取的最后一步,这些实体和关系的实例成为图中的单个节点和边。实体描述被聚合,并针对每个节点和边进行摘要。关系被聚合为图中的边,而同一关系的重复出现次数成为边的权重。断言也以类似方式聚合。
In this manuscript, our analysis uses exact string matching for entity matching – the task of reconciling different extracted names for the same entity ( Barlaug and Gulla,, 2021 ; Christen and Christen,, 2012 ; Elmagarmid et al.,, 2006 ) . However, softer matching approaches can be used with minor adjustments to prompts or code. Furthermore, GraphRAG is generally resilient to duplicate entities since duplicates are typically clustered together for summarization in subsequent steps.
在本文的分析中,实体匹配采用精确字符串匹配;实体匹配指的是协调同一实体被抽取出的不同名称这一任务(Barlaug 与 Gulla,2021;Christen 与 Christen,2012;Elmagarmid 等,2006)。不过,对提示词或代码作少量调整,就可以采用更宽松的匹配方法。此外,GraphRAG 通常能容忍重复实体,因为后续步骤往往会将它们聚在一起进行摘要。
本节正文与示例对照,图表及其他章节请查阅原文。译文尚未经人工专业审校。
一条错误新闻转载十次,可以当作十份独立证据吗?
划选正文也可以保存重点。笔记默认仅自己可见,确认公开后才进入共学社区。
重复抽取次数可以作为图算法的权重,但不等于事实正确的概率。能容忍重复实体,是作者在摘要任务中的判断,不保证业务身份合并正确。