# US 2021/0342716 A1 · Device and method for determining a knowledge graph

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0342716 A1
- Document ID: US20210342716A1
- Status: Published application
- Kind: A1
- Application number: 17/237,669
- Filed: 2021-04-22
- Published: 2021-11-04
- Earliest claimed: 2020-04-30
- Assignees: Robert Bosch GmbH
- Applicants: Robert Bosch GmbH
- Inventors: Heike Adel-Vu; Jannik Stroetgen; Lukas Lange
- Web page: https://patentgrep.com/patent/US20210342716A1

## CPC

- G06F 40/295: Named entity recognition
- G06N 5/04: Inference or reasoning models
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/0475: Generative networks
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06N 3/094: Adversarial learning
- G06N 5/022: Knowledge engineering; Knowledge acquisition
- G06N 20/00: Machine learning
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A device and method for determining a knowledge graph. A second embedding is determined for a first embedding for a word including a function. A first classification, which determines whether or not the word is an entity for the knowledge graph, or which defines to which entity or to which type of entity for the knowledge graph the word in the knowledge graph is to be assigned, is determined for the second embedding using a first classifier. A second classification, which defines to which type of embeddings from a plurality of types of embeddings the second embedding is to be assigned, is determined for the second embedding using a second classifier. At least one parameter for the function is trained in a training as a function of a gradient for the training of the first classifier and as a function of a gradient for the training of the second classifier.

## Claims

### Claim at position 1

1-15. (canceled)

### Claim 16

A method for determining a knowledge graph, the method comprising the following steps:
determining a second embedding for a first embedding for a word including a function;
determining for the second embedding, a first classification which defines whether or not the word is an entity for the knowledge graph, or which defines to which entity or to which type of entity for the knowledge graph the word in the knowledge graph, is to be assigned, using a first classifier;
determining, for the second embedding, a second classification which defines to which type of embeddings from a plurality of types of embeddings the second embedding is to be assigned, using a second classifier;
training at least one parameter for the function in a training as a function of at least one gradient for a training of the first classifier and as a function of at least one gradient for a training of the second classifier.

### Claim 17 (depends on 16)

The method as recited in claim 16, wherein a plurality of words is provided for the training, the first classification of the second embedding being determined using the first classifier for at least one portion of the plurality of words, the second classification of the second embedding being determined using the second classifier for the at least one portion of the plurality of words, and wherein parameters for the first classifier are determined as a function of least one portion of the plurality of words for which a first function for the first classifier fulfills a first condition, and wherein parameters for the second classifier are determined as a function of the parameters for the first classifier and as a function of at least one portion of the plurality of words for which a second function for the second classifier fulfills a second condition, the first function being defined as a function of the parameters for the first classifier and of the first classification, the second function being defined as a function of the parameters for the second classifier and of the second classification.

### Claim 18 (depends on 17)

The method as recited in claim 17, wherein parameters of the function are determined as a function of a difference between gradients for the first classifier and gradients for the second classifier.

### Claim 19 (depends on 16)

The method as recited in claim 16, wherein a plurality of second embeddings is determined, a group including various first embeddings being determined for each word from a plurality of different words and the group being mapped onto a group of second embeddings.

### Claim 20 (depends on 16)

The method as recited in claim 16, wherein various types of first embeddings are determined for the word, the various types of first embeddings encompassing a word-based embedding, and/or a letter-based embedding, and/or a subword-based embedding and/or a domain-specific embedding for a biomedical or materials science domain.

### Claim 21 (depends on 16)

The method as recited in claim 16, wherein the function maps at least one first embedding from a plurality of first embeddings onto a second embedding of a predefined dimension using an image.

### Claim 22 (depends on 21)

The method as recited in claim 21, wherein a weight for a combination of the second embedding is determined using at least one other second embedding as a function of the second embedding, the combination being determined as a function of the second embedding weighted with the weight.

### Claim 23 (depends on 22)

The method as recited in claim 22, wherein the combination is determined as a function of a product of the second embedding with the weight.

### Claim 24 (depends on 23)

The method as recited in claim 23, wherein a plurality of second embeddings are determined for the word, a plurality of products are determined for the plurality of second embeddings, and the combination is determined as a function of a sum of the products.

### Claim 25 (depends on 22)

The method as recited in claim 22, wherein the weight is determined as a function of a feature, which characterizes a property of the word.

### Claim 26 (depends on 22)

The method as recited in claim 22, wherein the weight is mapped using a softmax function onto values between 0 and 1, wherein weights for different embeddings of a predefined dimension for the word add up to 1.

### Claim 27 (depends on 16)

The method as recited in claim 16, wherein a vector defines an embedding, different vectors for various first embeddings for the word being mapped onto vectors of the same dimension using an image.

### Claim 28 (depends on 16)

The method as recited in claim 16, wherein the function of the second embedding is determined using a generator, which is defined by the at least one parameter for the function.

### Claim 29

A device for determining a knowledge graph, the device configured to:
determine a second embedding for a first embedding for a word including a function;
determine for the second embedding, a first classification which defines whether or not the word is an entity for the knowledge graph, or which defines to which entity or to which type of entity for the knowledge graph the word in the knowledge graph, is to be assigned, using a first classifier;
determine, for the second embedding, a second classification which defines to which type of embeddings from a plurality of types of embeddings the second embedding is to be assigned, using a second classifier;
train at least one parameter for the function in a training as a function of at least one gradient for a training of the first classifier and as a function of at least one gradient for a training of the second classifier.

### Claim 30

A non-transitory computer-readable medium on which is stored a computer program including computer-readable instructions, the instructions, when executed by a computer, causing the computer to perform the following steps:
determining a second embedding for a first embedding for a word including a function;
determining for the second embedding, a first classification which defines whether or not the word is an entity for the knowledge graph, or which defines to which entity or to which type of entity for the knowledge graph the word in the knowledge graph, is to be assigned, using a first classifier;
determining, for the second embedding, a second classification which defines to which type of embeddings from a plurality of types of embeddings the second embedding is to be assigned, using a second classifier;
training at least one parameter for the function in a training as a function of at least one gradient for a training of the first classifier and as a function of at least one gradient for a training of the second classifier.
