Methods and systems for approximating embeddings of out-of-knowledge-graph entities for link prediction in knowledge graph

Patent metadata

Field	Value
Publication number	US-12505358-B2
Application number	US-202117514512-A
Country	US
Kind code	B2
Filing date	Oct 29, 2021
Priority date	Oct 29, 2021
Publication date	Dec 23, 2025
Grant date	Dec 23, 2025

Abstract

The present disclosure describes methods and systems for generating an approximated embedding of an out-of-knowledge-graph entity based on a knowledge graph. The method includes: receiving a target entity, a dataset associated with the target entity, and an embeddings space of a knowledge graph comprising a set of structured data, wherein the target entity is out of the knowledge graph and the embeddings space includes a set of vectors representing the set of structured data in the embeddings space; selecting a set of elements from the knowledge graph, each element being related to the target entity according to the dataset associated with the target entity; constructing a set of descriptory triples based on the target entity and the set of elements; obtaining an embedding matrix based on the descriptory triples and the embeddings space; and generating an approximated embedding for the target entity based on the embedding matrix.

First claim

What is claimed is:

1 . A computing device for generating an approximated embedding of an out-of-knowledge-graph entity based on a knowledge graph, the computing device comprising:
a reception circuitry configured to receive a target entity, a dataset associated with the target entity, and an embeddings space of a knowledge graph comprising a set of structured data, wherein the set of structured data comprises a set of entities and links between the set of entities in the knowledge graph, wherein the target entity is out of the knowledge graph and the embeddings space includes a set of vectors representing the set of structured data associated with the knowledge graph in the embeddings space, and wherein the dataset associated with the target entity comprises data records of the target entity in triples stored in the knowledge graph; a selection circuitry configured to select a set of elements from the knowledge graph, each element being related to the target entity according to the dataset associated with the target entity; a construction circuitry configured to construct a set of descriptory triples based on the target entity and the set of elements, wherein the set of descriptory triples comprises the data records of the target entity, wherein each descriptory triple of the set of descriptory triples comprises the target entity, an entity selected from the set of elements, and a predicate obtained from the dataset associated with the target entity, and wherein the predicate defines a relational link between the target entity and the entity selected from the set of elements; an aggregation circuitry configured to obtain an embedding matrix based on the set of descriptory triples and the embeddings space, wherein the embedding matrix comprises a plurality of row vectors, wherein each row vector corresponds to a descriptory triple of the set of descriptory triples, wherein each row vector is formed by combining embeddings, retrieved from the embeddings space, of the target entity, the entity selected from the set of elements, and the predicate of the descriptory triple; an approximator circuitry configured to generate an approximated embedding for the target entity based on a pretrained embedding estimation model to the embedding matrix, wherein the approximated embedding indicates a vector representation corresponding to the target entity in the embeddings space of the knowledge graph; and an output circuitry configured to output a plausibility prediction of linkage between the target entity and one or more entities of the set of entities in the knowledge graph based on the approximated embedding for the target entity, wherein the plausibility prediction allows automatic integration of the target entity into the knowledge graph without retraining the embedding estimation model.

2 . The computing device according to claim 1 , wherein:
the set of elements from the knowledge graph comprises a set of neighbor relations and a set of neighbor entities according to the dataset associated with the target entity; and each of the descriptory triples comprises one element from the set of neighbor relations, one element from the set of neighbor entities, and the target entity.

3 . The computing device according to claim 1 , wherein:
the set of descriptory triples comprises a set of subjective descriptory triples and a set of objective descriptory triples; each triple of the set of subjective descriptory triples comprises the target entity as a subject in the triple and a subject along with a link belonging to the knowledge graph with known embeddings; and each triple of the set of objective descriptory triples comprises the target entity as the object in the triple and the subject along with a link belonging to the knowledge graph with known embeddings.

4 . The computing device according to claim 3 , wherein:
the embedding matrix comprises a first set of rows and a second set of rows; each row of the first set of rows corresponds to each triple of the set of subjective descriptory triples and comprises an embedding of the predicate of each triple of the set of subjective descriptory triples and an embedding of the object of each triple of the set of subjective descriptory triples; and each row of the second set of rows corresponds to each triple of the set of objective descriptory triples and comprises an embedding of the subject of each triple of the set of objective descriptory triples and an embedding of the predicate of each triple of the set of objective descriptory triples.

5 . The computing device according to claim 1 , wherein:
the embedding matrix comprising a matrix with m rows and n columns, m and n being positive integers; and a value of n is twice of a number of embedding dimensions of the embeddings space of the knowledge graph.

6 . The computing device according to claim 1 , wherein:
the embedding estimation model is pretrained by: for each sample entity or a sample subset of entities in the knowledge graph:
generating a modified embeddings space for the knowledge graph with the sample entity or the sample subset of entities excluded,
identifying neighborhood elements of the sample entity or the sample subset of entities in the knowledge graph,
generating a neighborhood embedding matrix according to the modified embeddings space and the neighborhood elements,
processing the neighborhood embedding matrix using the pretrained embedding estimation model to generate an approximated embedding for the sample entity or the sample subset of entities,
obtaining an estimation error based on the approximated embedding and a true embedding of the sample entity or the sample subset of entities in the embeddings space of the knowledge graph; and
adjusting a set of model parameters of the pretrained embedding estimation model by minimizing an overall error aggregated over the estimation error of each sample entity or the sample subset of entities in the knowledge graph.

7 . A method for generating an approximated embedding of one or more out-of-knowledge-graph entity based on a knowledge graph, the method comprising:
receiving, by a reception circuitry, a target entity, a dataset associated with the target entity, and an embeddings space of a knowledge graph comprising a set of structured data, wherein the set of structured data comprises a set of entities and links between the set of entities in the knowledge graph, wherein the target entity is out of the knowledge graph and the embeddings space includes a set of vectors representing the set of structured data associated with the knowledge graph in the embeddings space, and wherein the dataset associated with the target entity comprises data records of the target entity in triples stored in the knowledge graph; selecting, by a selection circuitry, a set of elements from the knowledge graph, each element being related to the target entity according to the dataset associated with the target entity; constructing, by a construction circuitry, a set of descriptory triples based on the target entity and the set of elements, wherein the set of descriptory triples comprises the data records of the target entity, wherein each descriptory triple of the set of descriptory triples comprises the target entity, an entity selected from the set of elements, and a predicate obtained from the dataset associated with the target entity, and wherein the predicate defines a relational link between the target entity and the entity selected from the set of elements; obtaining, by an aggregation circuitry, an embedding matrix based on the descriptory triples and the embeddings space, wherein the embedding matrix comprises a plurality of row vectors, wherein each row vector corresponds to a descriptory triple of the set

Assignees

• Accenture Global Solutions Ltd

Inventors

• Janik Adrianna
• Pai Sumit
• Costabello Luca

Classifications

• G06F18/2135

  based on approximation criteria, e.g. principal component analysis · CPC title

• G06N3/094

  Adversarial learning · CPC title

• G06N5/022 Primary

  Knowledge engineering; Knowledge acquisition · CPC title

• G06N5/02 Primary

  Knowledge representation; Symbolic representation · CPC title