Context-aware entity linking for knowledge graphs to support decision making

Patent metadata

Field	Value
Publication number	US-12596933-B2
Application number	US-202117392319-A
Country	US
Kind code	B2
Filing date	Aug 3, 2021
Priority date	Mar 19, 2021
Publication date	Apr 7, 2026
Grant date	Apr 7, 2026

Abstract

A machine learning model includes a context transformer and a decision head. The context transformer is a neural network of self-attention layers. The model makes a link prediction for a query embedding. Input embeddings are received at inputs of the context transformer. The input embeddings have: a query embedding set, the query embedding set comprising a subject embedding, object embedding, and relation embedding, one of the subject embedding, the object embedding, and the relation embedding being the query embedding; and knowledge graph embeddings. A first self-attention layer generates an attention score for each of the input embeddings. A final layer of the context transformer generates the link prediction for the query embedding and an output associated with each of the input embeddings. The decision head combines the attention score and the output for each of the input embeddings to determine a significance score for each of the input embeddings.

First claim

What is claimed is:

1 . A method for operating a machine learning model comprising a context transformer and a decision head, the context transformer comprising a neural network having a plurality of self-attention layers, the machine learning model configured to make a link prediction for a query embedding, the method comprising:
obtaining from a first data source of the one or more external data sources, one or more first knowledge graphs and first context data associated with the one or more first knowledge graphs; obtaining, from a second data source of the one or more external data sources, one or more second knowledge graphs and second context data associated with the one or more second knowledge graphs; canonicalizing the one or more first knowledge graphs, the first context data, the one or more second knowledge graphs, and the second context data to obtain canonicalized knowledge graphs and canonicalized context data; generating knowledge graph embeddings based on the canonicalized knowledge graphs and the canonicalized context data, receiving, input embeddings at inputs of the context transformer, the input embeddings comprising:
a query embedding set, the query embedding set comprising a subject embedding, object embedding, and relation embedding, wherein one of the subject embedding, the object embedding, and the relation embedding is the query embedding; and
the knowledge graph embeddings comprising first embeddings of triples associated with the respective knowledge graphs and second embeddings of the respective context data;
generating, by a first self-attention layer of the self-attention layers, an attention score for each of the input embeddings; generating, by a final layer of the neural network of the context transformer, the link prediction for the query embedding and an output associated with each of the input embeddings; combining, by the decision head, the attention score and the output for each of the input embeddings to determine a significance score for each of the input embeddings; and outputting the link prediction and the significance score for a subset of the input embeddings.

2 . The method of claim 1 ,
wherein the machine learning model is operating in a training phase, wherein, during the training phase, the query embedding set is a training embedding set having an associated ground truth vector for the query embedding, and wherein, during the training phase, the method further comprises:
outputting the link prediction as a link prediction vector;
comparing the link prediction vector with vectors of the knowledge graph embeddings; and
iteratively training the content transformer, using backpropagation, to change the link prediction vector until it is within a predetermined distance to the ground truth vector.

3 . The method of claim 2 ,
wherein the decision head comprises a multilayer perceptron with sigmoid activation functions, and wherein, during the training phase, the method further comprises:
receiving a predetermined significance score associated with at least one of the knowledge graph embeddings;
iteratively training the decision head, using backpropagation, to change the significance score generated for the at least one of the knowledge graph embeddings until it is within a second predetermined distance to the predetermined significance score.

4 . The method of claim 3 , wherein the predetermined significance score is provided by a human operator.

5 . The method of claim 2 , the method further comprising, prior to a training phase:
transforming the triples of the respective knowledge graphs into self-supervised training examples by masking an object, relation, or subject of each of the triples, each of the self-supervised training examples comprising an incomplete triple and associated ground truth data indicated the masked one of the object, relation or subject; receiving, from the one or more external data sources, the respective context data comprising a plurality of context relevant to the respective knowledge graphs; respectively associate individual context of the plurality of context with relevant ones of entities, relations, and the triples of the respective knowledge graph; determine for each of the entities, the relations, the triples, and the plurality of context a type and a group; generate the query embedding from a selected one of the self-supervised training examples and the determined type and group of the selected one of the self-supervised training examples; and generate the knowledge graph embeddings from the triples of the respective knowledge graphs and the type and the group for the entities, the relations, the triples, and the plurality of context.

6 . The method of claim 5 , wherein the method comprises splitting the self-supervised training examples into a training set and a validation set at random.

7 . The method of claim 6 , the method further comprising, after a training phase, a test phase comprising:
for a plurality of the self-supervised training examples in the validation set:
receiving, at query inputs of the inputs of the context transformer, a current validation embedding set, the current validation embedding set comprising a respective one of the plurality of the self-supervised training examples from the validation set;
generating, by the first self-attention layer of the self-attention layers, the attention score for each of the knowledge graph embeddings and the current validation embedding set,
generating, by the final layer of the neural network of the context transformer, the link prediction for the current validation embedding set and the output associated with each of the input embeddings,
combining, by the decision head, the attention score and the output for each of the input embeddings to determine the significance score for each of the input embeddings; and
outputting the link prediction and the significance score for the subset of the input embeddings; and
determining an accuracy of model based on comparing the link prediction and the associated ground truth data associated with each of the plurality the plurality of the self-supervised training examples from the validation set.

8 . The method of claim 1 , wherein the knowledge graph embeddings are non-monotonous, double positional embeddings.

9 . The method of claim 1 , wherein the knowledge graph embeddings further comprise a related type-embedding and a related coherence-embedding.

10 . The method according to claim 1 , wherein in an entity alignment mode, the method comprises providing the query embedding set with the relation embedding being a SameAs embedding and one of the subject embedding and the object embedding being the query embedding.

11 . The method according to claim 1 , the method comprising:
outputting the link prediction and the significance score via a human-machine interface; receiving feedback input from the human-machine interface; and and updating the at least one of the knowledge graphs according to a link prediction based on the feedback indicating an accepted link prediction.

12 . The method of claim 1 , wherein generating the knowledge graph embeddings comprises:
merging the canonicalized knowledge graphs and the canonicalized context data into a joint knowledge graph; seeding the joint knowledge graph with equivalent entities, wherein the equivalent entities are triples comprising a first, second, and third component, wherein the first and the third component are SameAs facts; and generating the knowledge graph embeddings based on the joint knowledge graph comprising the seeded equivalent entities.

13 .

Assignees

• NEC Laboratories Europe GmbH
• Nec Corp

Inventors

• Friede David
• Gashteovski Kiril

Classifications

• G06N3/08

  Learning methods · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/091

  Active learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title