Granted Patent B2

US 12,541,693 · App. 17/973,322 · Granted Feb 3, 2026

Inventors: Jan Portisch (Bruchsal, DE); Sandra Bracholdt (Dielheim, DE); Michael Hoerisch (Heidelberg, DE)

Assignee: SAP SE

G06N5/022

Abstract

Claims (75)

1 . A knowledge graph computing system, comprising:

at least one processor programmed to perform operations comprising:

extracting first relationship data from a first document, the first relationship data describing a first subject entity, a first object, and a first predicate describing a relationship between the first subject entity and the first object;

determining a first confidence value describing the first relationship data;

writing the first relationship data to a knowledge graph data structure stored at a data store, the first relationship data being written to the knowledge graph data structure as a first triple data unit comprising:

a first subject subunit describing the first subject entity;

a first object subunit describing the first object;

a first predicate subunit describing the first predicate; and

a first confidence subunit describing the first confidence value;

serving a first user interface page to a user computing device associated with a first user, the first user interface page comprising an indication of the first subject entity, the first predicate, and the first object;

receiving, via the first user interface page, feedback data describing an accuracy of the first relationship data; and

modifying the first confidence subunit to describe an updated confidence value, the updated confidence value being based on the feedback data and a trust score of the first user.

2 . The system of claim 1 , the operations further comprising:

generating training data based at least in part on the feedback data; and

training a machine learning model to extract relationships from documents, the training being based at least in part on the training data.

3 . The system of claim 2 , the extracting of the first relationship data from the first document being performed using the machine learning model, and the training of the machine learning model being a re-training of the machine learning model.

4 . The system of claim 1 , the first triple data unit further comprising a confidentiality data unit describing a confidentiality level of the first relationship data, the operations further comprising:

receiving a query against the knowledge graph data structure;

determining that the first relationship data is responsive to the query; and

determining that the confidentiality level of the first relationship data matches a confidentiality level associated with the first user.

5 . The system of claim 1 , the determining of the first confidence value being based at least in part on a trust score associated with an author of the first document.

6 . The system of claim 1 , the operations further comprising:

accessing, from a sense library at the data store, first sense data describing a first sense of the first subject entity and second sense data describing a second sense of the first subject entity; and

determining that the first relationship data indicates the first sense of the first subject entity, the first triple data unit further comprising a first sense subunit describing the first sense.

7 . The system of claim 1 , the operations further comprising, before writing the first relationship data to the knowledge graph data structure, determining, by the computing system, the first relationship data is not represented at the knowledge graph data structure.

8 . The system of claim 1 , the operations further comprising:

accessing second relationship data describing a second subject entity, a second object, and a second predicate describing relationship between the second subject entity and the first object;

determining a second confidence value describing the second relationship data;

determining that the second relationship data is represented at the knowledge graph data structure at a second triple data unit; and

modifying a second confidence subunit of the second triple data unit based at least in part on the second confidence value.

9 . The system of claim 1 , the extracting being performed by a relationship extractor system, determining of the first confidence value describing the first relationship data being based at least in part on, a confidence value of the relationship extractor system and a trust score associated with an author of the first document.

10 . A method for utilizing relationship data in a computing system, the method comprising:

extracting, by the computing system, first relationship data from a first document, the first relationship data describing a first subject entity, a first object, and a first predicate describing a relationship between the first subject entity and the first object;

determining, by the computing system, a first confidence value describing the first relationship data;

writing, by the computing system, the first relationship data to a knowledge graph data structure stored at a data store associated with the computing system, the first relationship data being written to the knowledge graph data structure as a first triple data unit comprising:

a first subject subunit describing the first subject entity;

a first object subunit describing the first object;

a first predicate subunit describing the first predicate; and

a first confidence subunit describing the first confidence value;

serving, by the computing system, a first user interface page to a user computing device associated with a first user, the first user interface page comprising an indication of the first subject entity, the first predicate, and the first object;

receiving, by the computing system and via the first user interface page, feedback data describing an accuracy of the first relationship data; and

modifying, by the computing system, the first confidence subunit to describe an updated confidence value, the updated confidence value being based on the feedback data and a trust score of the first user.

11 . The method of claim 10 , further comprising:

generating training data based at least in part on the feedback data; and

training a machine learning model to extract relationships from documents, the training being based at least in part on the training data.

12 . The method of claim 11 , the extracting of the first relationship data from the first document being performed using the machine learning model, and the training of the machine learning model being a re-training of the machine learning model.

13 . The method of claim 10 , the first triple data unit further comprising a confidentiality data unit describing a confidentiality level of the first relationship data, the method further comprising:

receiving, by the computing system, a query against the knowledge graph data structure;

determining, by the computing system, that the first relationship data is responsive to the query; and

determining that the confidentiality level of the first relationship data matches a confidentiality level associated with the first user.

14 . The method of claim 10 , the determining of the first confidence value being based at least in part on a trust score associated with an author of the first document.

15 . The method of claim 10 , further comprising:

accessing, from a sense library at the data store, first sense data describing a first sense of the first subject entity and second sense data describing a second sense of the first subject entity; and

determining, by the computing system, that the first relationship data indicates the first sense of the first subject entity, the first triple data unit further comprising a first sense subunit describing the first sense.

16 . The method of claim 10 , further comprising, before writing the first relationship data to the knowledge graph data structure, determining, by the computing system, the first relationship data is not represented at the knowledge graph data structure.

17 . The method of claim 10 , further comprising:

accessing, by the computing system, second relationship data describing a second subject entity, a second object, and a second predicate describing relationship between the second subject entity and the first object;

determining, by the computing system, a second confidence value describing the second relationship data;

determining, by the computing system, that the second relationship data is represented at the knowledge graph data structure at a second triple data unit; and

modifying a second confidence subunit of the second triple data unit based at least in part on the second confidence value.

18 . The method of claim 10 , the extracting being performed by a relationship extractor system, determining of the first confidence value describing the first relationship data being based at least in part on, a confidence value of the relationship extractor system and a trust score associated with an author of the first document.

19 . A machine-readable medium comprising instructions stored thereon that, when executed by at least one processor, cause the at least one processor to perform operations comprising:

extracting first relationship data from a first document, the first relationship data describing a first subject entity, a first object, and a first predicate describing a relationship between the first subject entity and the first object;

determining a first confidence value describing the first relationship data;

writing the first relationship data to a knowledge graph data structure stored at a data store, the first relationship data being written to the knowledge graph data structure as a first triple data unit comprising:

a first subject subunit describing the first subject entity;

a first object subunit describing the first object;

a first predicate subunit describing the first predicate; and

a first confidence subunit describing the first confidence value;

serving a first user interface page to a user computing device associated with a first user, the first user interface page comprising an indication of the first subject entity, the first predicate, and the first object;

receiving, via the first user interface page, feedback data describing an accuracy of the first relationship data; and

modifying the first confidence subunit to describe an updated confidence value, the updated confidence value being based on the feedback data and a trust score of the first user.

20 . The medium of claim 19 , the operations further comprising:

generating training data based at least in part on the feedback data; and

training a machine learning model to extract relationships from documents, the training being based at least in part on the training data.

Assignments (2)

Continuity (2)

References Cited (26)