Predicate redundancy in the Open Research Knowledge Graph

John, Aggarwal, Auer & Karras (TIB Hannover / Leibniz University Hannover), Addressing Predicate Redundancy in Research Knowledge Graphs: Duplicate Detection, Resolution, and Prevention, TPDL 2026 (arXiv:2607.03197v2). Read in full. A duplicate-rate measurement on a live, crowdsourced scholarly knowledge graph, the Open Research Knowledge Graph. It is the first source here that measures redundancy in the relations rather than the entities.

The number

Clustering ORKG's 13,457 predicates into 3,837 clusters gives 30% potentially removable, split into two different problems:

Behind the headline: 2,044 predicates (15.19%) across 786 clusters are redundant merge candidates that could collapse to one representative each, and 1,989 (14.78%) are simply unused. The cluster-size distribution is severely skewed — 2,769 singletons against a handful of clusters between 100 and 477 members.

Why the duplicates exist

The paper's real contribution is that it went and looked, and the causes are behavioural rather than algorithmic:

• The same user creating the same predicate repeatedly, often minutes apart. 593 clusters contain predicates created by a single user, and 79 clusters contain identical labels. ORKG lets you create a predicate whose label already exists, with no warning.
• IDs typed where labels belong. Searching by identifier requires a leading #, which people miss, so raw IDs land in label fields and produce lexical similarity between unrelated predicates. The 312-member cluster of IDs is a false positive of the detector, caused by an interface affordance.
• No standardisation for the common cases — several users independently inventing predicates for method, data, material. These are not always true duplicates, and the paper says resolving them needs human judgement and usage context.

Where this leaves the corpus's own question

This wiki has held entity-resolution as the problem of deciding when two nodes are the same thing, with core-kg and link-kg supplying the numbers (GraphRAG's 27–36% duplicate nodes falling to 10.6–17.8%). Those counts come from matching people and places, where a string comparison means something. The open question since 2026-08-08 has been what happens when the units are claim-shaped and no matcher can compare them.

Predicates are the nearest measured case. A relation like has evaluation against evaluated on is not resolvable by surface form, and the paper's numbers show what that costs at scale: 30% of the vocabulary is removable, and more than half the predicates land in Review — clusters where the method declines to recommend anything and a human has to decide. The authors are explicit that computed similarity does not imply true duplication, and that the clustering underestimates redundancy whenever semantically related predicates land in different clusters.

So the answer forming is that at claim level there is no threshold that works. The detector's job changes from deciding to triaging, and the largest bucket is the one it hands back. That is the same shape er-evaluation-framework describes for entity resolution evaluation, arrived at from the data-quality side rather than the metric side.

What it does not settle

Predicates are a graph's schema, not its content. A predicate vocabulary is small (13k), reused constantly, and visible in an interface where a curator can inspect it; a body of extracted propositions is large, mostly used once, and nobody looks at it. The duplicate rate on proposition-shaped nodes is still unmeasured, and this paper's method — embedding, cluster, hand-validate — does not obviously survive the scale change.