---
type: "post"
id: "knowledge-graphs-drift-out-of-date-as-organizations-b8e6b6"
slug: "knowledge-graphs-drift-out-of-date-as-organizations-b8e6b6"
title: "Knowledge graphs drift out of date as organizations merge, people change jobs, new entities appear, and old relationships expire."
description: "EMERGE is a benchmark for updating existing knowledge graphs from new text, rather than extracting facts without regard to what a graph already contains."
retrieval_nugget: "EMERGE is a benchmark for updating existing knowledge graphs from new text, rather than extracting facts without regard to what a graph already contains. Built from Wikipedia passages and Wikidata snapshots, it frames stale facts, structural integration and removal of outdated relations as measurable tasks for knowledge-system builders."
published_at: "2026-11-11"
updated_at: "2026-11-11"
record_date: "2026-11-11"
date_kind: "published_at"
topics: []
entities: ["EMERGE","Wikidata","Wikipedia","Klim Zaporojets","Daniel Daza","Edoardo Barba","Ira Assent","Roberto Navigli"]
related_patterns: ["company-memory-needs-write-loops"]
novelty_label: "new_fact"
what_changed: "EMERGE adds a benchmark for updating knowledge graphs from evolving text, pairing 233,000 Wikipedia passages with 1.19 million Wikidata edits across seven yearly snapshots."
prior_refs: ["https://newruntime.com/posts/google-cloud-knowledge-catalog-adds-ga-ingestion-of-4d9b4a/","https://newruntime.com/posts/sarvam-s-knowledge-fabric-compiles-enterprise-docs-into-4ffcbd/"]
source_url: "https://arxiv.org/abs/2507.03617"
source_urls: ["https://arxiv.org/abs/2507.03617"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:knowledge-graphs-drift-out-of-date-as-organizations-b8e6b6"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/knowledge-graphs-drift-out-of-date-as-organizations-b8e6b6/","markdown":"https://newruntime.com/posts/knowledge-graphs-drift-out-of-date-as-organizations-b8e6b6.md","json":"https://newruntime.com/posts/knowledge-graphs-drift-out-of-date-as-organizations-b8e6b6.json"}
markdown_url: "https://newruntime.com/posts/knowledge-graphs-drift-out-of-date-as-organizations-b8e6b6.md"
---

# Knowledge graphs drift out of date as organizations merge, people change jobs, new entities appear, and old relationships expire.

## Retrieval answer

EMERGE is a benchmark for updating existing knowledge graphs from new text, rather than extracting facts without regard to what a graph already contains. Built from Wikipedia passages and Wikidata snapshots, it frames stale facts, structural integration and removal of outdated relations as measurable tasks for knowledge-system builders.

## Summary
EMERGE is a benchmark for updating existing knowledge graphs from new text, rather than extracting facts without regard to what a graph already contains. Built from Wikipedia passages and Wikidata snapshots, it frames stale facts, structural integration and removal of outdated relations as measurable tasks for knowledge-system builders.

## What happened
The paper “EMERGE: A Benchmark for Updating Knowledge Graphs with Emerging Textual Knowledge” studies how to automatically update a knowledge graph over time using information in unstructured text. Its authors are Klim Zaporojets, Daniel Daza, Edoardo Barba, Ira Assent, Roberto Navigli and Paul Groth. The arXiv record lists the first submission on July 4, 2025, and a revised version, v3, dated November 7, 2026.

The paper distinguishes this from conventional information-extraction pipelines, which extract knowledge from text independently of the current state of a graph. For an update task, a system must identify the edit operations induced by a passage against a particular graph snapshot. That requires it to connect text to entities and relations that already exist, rather than treating every extracted statement as a standalone triple. It must also handle cases in which incoming information changes the status of a previously represented relation.

The authors propose a generic, extensible pipeline that pairs textual passages with the knowledge-graph edit operations they induce on a given snapshot. They instantiate it with the Wikidata knowledge graph and the English Wikipedia corpus. The resulting EMERGE dataset contains 233,000 Wikipedia passages associated with 1.19 million graph edits across seven yearly Wikidata snapshots, covering 2019 through 2025.

The paper presents EMERGE as a benchmark for methods intended to keep graphs current as knowledge emerges in text. Its experiments highlight challenges in integrating textual knowledge with an existing graph structure. The paper further reports that access to the graph substantially improves executable updates and enables triple deprecation: marking a represented relation as no longer valid. This means the benchmark includes more than the addition of new relations; it is designed around editing a changing representation of knowledge.

According to the arXiv record, the dataset and code are publicly available. EMERGE is not presented as a hosted product, commercial service or API. The abstract gives no pricing, service limits, deployment requirements or model-by-model score table. Its stated contribution is a dataset, an update-oriented task framing and experimental findings about using current graph state.

## What the archive adds
New Runtime’s earlier coverage described [Sarvam’s knowledge fabric](https://newruntime.com/posts/sarvam-s-knowledge-fabric-compiles-enterprise-docs-into-4ffcbd/) on October 26, 2026, and [Google Cloud Knowledge Catalog’s semantic-model ingestion](https://newruntime.com/posts/google-cloud-knowledge-catalog-adds-ga-ingestion-of-4d9b4a/) on October 31, 2026. Those were product developments in organizing enterprise knowledge; EMERGE is a new research event that makes one maintenance problem—updating a graph as evidence changes—more explicit and testable.

## Trend signal
EMERGE supports the [company-memory write-loops pattern](https://newruntime.com/patterns/company-memory-needs-write-loops/): useful organizational memory requires ongoing updates and correction, not just retrieval from a larger corpus. The benchmark treats new text as a reason to edit existing knowledge, including deprecating relations. It does not establish that a production system can safely automate that process, but it offers researchers a structured way to test parts of it.

## Why it matters
A knowledge system can retrieve a relevant passage and still mislead if its stored graph preserves an obsolete relation or fails to reconcile new information with what it already knows. EMERGE draws attention to that maintenance burden. For teams building retrieval-augmented applications, internal assistants or entity-centric data products, correctness depends not only on ingesting evidence but also on deciding how it changes the current representation.

The benchmark’s emphasis on graph access is operationally relevant. A pipeline that sees only new text can emit isolated facts without testing whether they fit the entities and relations already stored. A pipeline that can inspect the graph has the information needed to formulate updates in context, including potential removals. Neither approach alone resolves entity ambiguity, conflicting sources or authorization over who may change shared knowledge.

Teams can use EMERGE’s task definition as an evaluation prompt. Ask whether a system is tested on additions, corrections and removals, not only whether newly extracted triples appear plausible. Also separate a model’s ability to propose an edit from the surrounding system’s ability to validate, attribute and apply it. The latter concerns become more important as a graph is reused by multiple applications and agents.

## Limits and uncertainty
The paper’s findings concern benchmark construction and experiments; they do not show that EMERGE represents every enterprise graph or domain. The arXiv abstract does not establish performance on private, multilingual or rapidly changing organizational data. It also does not specify review workflows for proposed changes, permissions for applying edits, or the frequency at which a live graph should be refreshed.

The reported improvement from graph access is not accompanied in the abstract by effect sizes or a breakdown by update type. The available description also does not demonstrate reliability when text conflicts with an existing graph or when the graph itself contains an earlier error. Independent replication and production evidence would be needed before benchmark performance could be treated as proof of reliable live knowledge maintenance.

## Skill to build now
This week, add an update test set to one knowledge workflow. Select a small group of entities with known changes over time, then test whether the system can propose an addition, a correction and a retirement while showing the prior graph state that informed each proposal. Keep proposed edits separate from automatic application until the team has defined validation and ownership rules.

## Reader takeaway
EMERGE makes a useful distinction: keeping a knowledge graph current is not the same task as extracting facts from fresh text. If your system depends on stored entities and relationships, evaluate whether it can reconcile new evidence with existing state—and retire outdated facts—not just add more triples.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)