What is knowledge graph extraction?
Knowledge graph extraction reads documents and identifies the entities they mention and the relationships they state between them. The extracted facts are resolved against entities already in the graph, so that one team or product is represented once however many documents refer to it. Each node and edge is then stored with a reference to the passage it came from.
Also called: Graph extraction, Entity and relation extraction
Why a graph is only as reliable as its extraction
A graph index gives retrieval the ability to follow connections, and every connection it follows was created by extraction. An incorrect relationship is not an occasional bad answer. It becomes a permanent route that traversal will use confidently and cite as evidence. Extraction therefore needs the same discipline as any other data pipeline. Every fact should carry its source, and every relationship type should be defined in advance instead of invented per document.
- Every edge the graph contains was introduced by extraction.
- Sourced facts can be checked, where unsourced ones can only be trusted.
- A fixed schema of relationship types keeps the graph queryable.
- Resolution prevents one entity from appearing under several names.
How knowledge graph extraction works
- A schema defines the entity types and relationship types to be extracted.
- A model reads each passage and proposes entities and relationships within that schema.
- Proposed entities are resolved against existing nodes using names and attributes.
- Each accepted fact is stored with the document version and passage it came from.
- Facts drawn from a removed or superseded document are retired with it.
- Low-confidence extractions are queued for review before entering the graph.
How knowledge graph extraction is evaluated
Knowledge graph extraction is evaluated against a manually annotated sample of documents. Measure the following against those annotations:
- precision of extracted relationships against the annotations
- recall of relationships the annotators identified
- duplicate entities created for an existing node
- share of facts carrying a valid source reference
- relationships extracted outside the defined schema
Limitations, risks and common failure modes
- The model infers relationships a document implies but never states.
- Resolution merges distinct entities that happen to share a name.
- Facts from superseded documents remain in the graph after the update.
- Schema drift creates near-identical relationship types that queries treat separately.