Determining explanations for predicted links in knowledge graphs

Patent metadata

Field	Value
Publication number	US-10877979-B2
Application number	US-201815940298-A
Country	US
Kind code	B2
Filing date	Mar 29, 2018
Priority date	Jan 16, 2018
Publication date	Dec 29, 2020
Grant date	Dec 29, 2020

Abstract

A device receives a knowledge graph and an ontology for the knowledge graph, and receives a query for information associated with the knowledge graph. The device generates candidate responses to the query, and assigns scores to the candidate responses based on the knowledge graph. The device identifies a particular candidate response, of the candidate responses, based on the scores for the candidate responses, and determines, based on the knowledge graph, a neighborhood of the particular candidate response. The device generates knowledge graph embeddings for the neighborhood of the particular candidate response, and determines a particular neighborhood, with a smallest loss of quality, based on the knowledge graph embeddings. The device generates a reasoning graph based on the ontology and the particular neighborhood, and generates an explanation of the particular candidate response based on the reasoning graph. The device performs an action based the explanation of the particular candidate response.

First claim

What is claimed is:

1. A device, comprising:
one or more memories; and
one or more processors, communicatively coupled to the one or more memories, to:
receive a knowledge graph and an ontology for the knowledge graph;
receive a query for information associated with the knowledge graph;
generate candidate responses to the query based on the knowledge graph;
assign scores to the candidate responses based on the knowledge graph;
identify a particular candidate response, of the candidate responses, based on the scores for the candidate responses;
determine, based on the knowledge graph, a neighborhood of the particular candidate response;
generate knowledge graph embeddings for the neighborhood of the particular candidate response;
determine a subset of the neighborhood, with a smallest loss of quality, based on the knowledge graph embeddings;
generate a reasoning graph based on the ontology and the subset of the neighborhood,
wherein the reasoning graph includes a first level of abstraction and a second level of abstraction,
wherein the first level of abstraction and the second level of abstraction are determined based on a degree of abstraction that increases when moving in a particular direction through the reasoning graph;

generate an explanation of and an achieved score associated with the particular candidate response based on the reasoning graph; and
perform an action based the explanation of the particular candidate response.

2. The device of claim 1 , wherein the one or more processors are further to:
receive an embedding with a predicted link, and
wherein the one or more processors, when determining the subset of the neighborhood with the smallest loss of quality, are to:
compare a quality of the knowledge graph embeddings for the neighborhood relative to the received embedding; and
determine the subset of the neighborhood with the smallest loss of quality based on comparing the quality of the knowledge graph embeddings.

3. The device of claim 1 , wherein the one or more processors, when performing the action, are to:
provide information identifying the particular candidate response and the explanation of the particular candidate response.

4. The device of claim 1 , wherein the one or more processors, when determining the neighborhood of the particular candidate response, are to:
select a neighborhood sampling technique, from a plurality of neighborhood sampling techniques, to determine the neighborhood of the particular candidate response,
wherein the plurality of neighborhood sampling techniques includes:
an exhaustive technique,
a random walk technique,
a graph-traversal technique,
a degree-based technique, and
an evolutionary technique.

5. The device of claim 1 , wherein the one or more processors, when determining the subset of the neighborhood with the smallest loss of quality, are to:
utilize a loss of quality computation to determine the subset of the neighborhood with the smallest loss of quality,
wherein the loss of quality computation includes one or more of:
a Kruskal stress calculation,
a Sammon stress calculation,
a residual variance calculation,
a relative error calculation, or
a normalization independent embedding quality assessment (NIEQA) calculation.

6. The device of claim 1 , wherein the explanation of the particular candidate response includes two or more different levels of abstraction associated with the explanation.

7. The device of claim 1 , wherein the one or more processors, when determining the subset of the neighborhood with the smallest loss of quality, are to:
select a neighborhood sampling technique, from a plurality of neighborhood sampling techniques; and
determine the subset of the neighborhood with the smallest loss of quality based on the selected neighborhood sampling technique.

8. A non-transitory computer-readable medium storing instructions, the instructions comprising:
one or more instructions that, when executed by one or more processors, cause the one or more processors to:
receive a knowledge graph generated based on training data and an ontology for the training data,
the training data including information associated with a subject of the ontology;

receive a query for information associated with the knowledge graph;
generate candidate responses to the query based on the knowledge graph;
identify a particular candidate response, of the candidate responses, based on scoring the candidate responses based on the knowledge graph;
determine, based on the knowledge graph, a neighborhood of the particular candidate response;
generate knowledge graph embeddings for the neighborhood of the particular candidate response;
identify, based on the knowledge graph embeddings, a portion of the neighborhood with a smallest loss of quality;
generate a reasoning graph based on the ontology and the portion of the neighborhood,
wherein the reasoning graph includes a first level of abstraction and a second level of abstraction,
wherein the first level of abstraction and the second level of abstraction are determined based on a degree of abstraction that increases when moving in a particular direction through the reasoning graph;

generate an explanation of and an achieved score associated with the particular candidate response based on the reasoning graph; and
perform one or more actions based the explanation of the particular candidate response.

9. The non-transitory computer-readable medium of claim 8 , wherein the one or more instructions, that cause the one or more processors to perform the one or more actions, cause the one or more processors to:
provide, for display, information identifying the particular candidate response and the explanation of the particular candidate response.

10. The non-transitory computer-readable medium of claim 8 , wherein the one or more instructions, that cause the one or more processors to generate the reasoning graph, cause the one or more processors to:
process the ontology and the portion of the neighborhood, with a reasoning model, to generate the reasoning graph,
wherein the reasoning model includes one or more of:
a resource description framework (RDF) model,
a RDF schema (RDFS) model, or
a web ontology language (OWL) model.

11. The non-transitory computer-readable medium of claim 8 , wherein the one or more instructions, that cause the one or more processors to determine the neighborhood of the particular candidate response, cause the one or more processors to:
utilize one or more neighborhood sampling techniques to determine the neighborhood of the particular candidate response.

12. The non-transitory computer-readable medium of claim 8 , wherein the one or more instructions, that cause the one or more processors to identify the portion of the neighborhood with the smallest loss of quality, cause the one or more processors to:
utilize a loss of quality computation to identify the portion of the neighborhood with the smallest loss of quality.

13. The non-transitory computer-readable medium of claim 8 , wherein the explanation of the particular candidate response includes two or more different levels of abstraction associated with the explanation.

14. The non-transitory computer-readable medium of claim 8 , wherein the instructions further comprise:
one or more instructions that, when executed by the one or more processors, cause the one or more processors to:
utilize a relational learning model to determine values associated with the candidate responses; and
utilize the values to score the candidate responses.

15. A method, compris

Assignees

• Accenture Global Solutions Ltd

Inventors

• Costabello Luca
• Lecue Freddy

Classifications

• G06F16/9024 Primary

  Graphs; Linked lists (G06F16/9027 takes precedence) · CPC title

• G16H70/00

  ICT specially adapted for the handling or processing of medical references · CPC title

• G06N5/02

  Knowledge representation; Symbolic representation · CPC title

• G16H50/70

  for mining of medical data, e.g. analysing previous cases of other patients · CPC title

• G06N5/022

  Knowledge engineering; Knowledge acquisition · CPC title