# US 2021/0279512 A1 · Computer-implemented method and device for text analysis > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 2021/0279512 A1 - Document ID: US20210279512A1 - Status: Published application - Kind: A1 - Application number: 17/259,864 - Filed: 2019-07-11 - Published: 2021-09-09 - Earliest claimed: 2018-08-03 - Assignees: Robert Bosch GmbH - Applicants: Robert Bosch GmbH - Inventors: Anna Constanze Haetty; Michael Dorna - Web page: https://patentgrep.com/patent/US20210279512A1 ## CPC - G06N 3/084: Backpropagation, e.g. using gradient descent - G06F 18/214: Generating training patterns; Bootstrap methods, e.g. bagging or boosting - G06F 18/2431: Multiple classes - G06N 3/042: Knowledge-based neural networks; Logical representations of neural networks - G06N 3/045: Combinations of networks - G06N 3/0499: Feedforward networks - G06N 3/08: Learning methods - G06N 3/09: Supervised learning - G06N 3/048: Activation functions - G06N 5/022: Knowledge engineering; Knowledge acquisition - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) ## Abstract A computer-implemented method for training an artificial neural network using training data, which include features and identifiers, the features characterizing term candidates from a corpus, the corpus including a text from a domain, the identifier characterizing a degree of association to at least three classes for term candidates that differ from one another, different classes indicating different degrees of association of the term candidates to the domain, the training data including an assignment of features to identifiers. An artificial neural network, method for classifying term candidates, and computer-implemented method for generating training data, are also described. ## Claims ### Claim at position 1 1-20. (canceled) ### Claim 21 A computer-implemented method for training an artificial neural network using training data, which include features and identifiers, the features characterizing term candidates from a corpus, the corpus encompassing a text from a domain, each of the identifiers characterizing a degree of association to at least three classes for the term candidates that differ from one another, the different classes indicating different degrees of association of the term candidate to the domain, the training data including an assignment of the features to the identifiers, the method comprising the following steps: predefining a feature of the features to an input layer of the artificial neural network; assigning, by the artificial neural network, an identifier of the identifiers to the feature from the input layer in a prediction in an output layer of the artificial neural network; comparing the identifier from the output layer with the identifier assigned to the feature in the training data; and learning, as a function of the result of the comparison, at least one parameter of the artificial neural network, which characterizes a connection of the artificial neural network between the input layer and the output layer. ### Claim 22 (depends on 21) The method as recited in claim 21, wherein the term candidates are drawn from a corpus subject-specific to the domain. ### Claim 23 (depends on 22) The method as recited in claim 22, wherein the term candidates are assigned to at least one of the classes and the features for the term candidates are determined, a word vector being determined, and at least one parameter of the artificial neural network being trained using the features for the term candidates. ### Claim 24 (depends on 21) The method as recited in claim 21, wherein the term candidates are compounds including at least two components. ### Claim 25 (depends on 24) The method as recited in claim 24, wherein the compounds are from a corpus subject-specific to the domain and are split as term candidates into components, the compounds are assigned to at least one of the classes, the features for the compounds and the components are determined, and at least one parameter of the artificial neural network is trained using the features for the compounds and the components. ### Claim 26 (depends on 25) The method as recited in claim 25, wherein at least one word vector is determined as at least one of the features. ### Claim 27 (depends on 25) The method as recited in claim 25, wherein a productivity and a frequency of the components are determined as the features for the compounds and the components based on the corpus. ### Claim 28 A computer-implemented method for generating training data for a training of an artificial neural network, the training data including features and identifiers, the method comprising the following steps: determining features which characterize term candidates from a corpus, the corpus including a text from a domain; determining a respective identifier which characterizes a degree of association to at least three classes for the term candidates that differ from one another, the different classes indicating different degrees of association of the term candidates to the domains; and assigning at least one of the identifiers to at least one of the features. ### Claim 29 (depends on 28) The method as recited in claim 28, wherein the term candidates are drawn from a corpus subject-specific to the domain. ### Claim 30 (depends on 29) The method as recited in claim 29, wherein the term candidates are assigned to at least one of the classes and the features are determined for the term candidates, a word vector being determined. ### Claim 31 (depends on 28) The method as recited in claim 28, wherein the term candidates are compounds including at least two components. ### Claim 32 (depends on 31) The method as recited in claim 31, wherein the compounds are from a corpus subject-specific to the domain and are split as term candidates into the components, the compounds are assigned to at least one of the classes, and the features for the compounds and the components are determined. ### Claim 33 (depends on 32) The method as recited in claim 32, wherein at least one word vector is determined as at least one of the features. ### Claim 34 (depends on 32) The method as recited in claim 32, wherein a productivity and a frequency of the components are determined as the features based on the corpus. ### Claim 35 An artificial neural network, comprising: an input layer, a feature being predefinable for the input layer; and an output layer, the artificial neural network being configured to assign an identifier to the feature from the input layer in a prediction in the output layer, the feature characterizing term candidates from a corpus, the corpus including a text from a domain, the identifier characterizing at least three classes for the term candidates that differ from one another, different classes indicating different degrees of association of the term candidates to the domain. ### Claim 36 (depends on 35) The artificial neural network as recited in claim 35, wherein the artificial neural network includes at least one first input layer to which a compound and its components are predefinable for a first feature, the artificial neural network including at least one second input layer for which a productivity and a frequency of the components are predefinable for a second feature, the output layer being situated downstream from the first and second input layers and outputting the identifier in the prediction as a function of the first feature and of the second feature. ### Claim 37 (depends on 36) The artificial neural network as recited in claim 36, wherein the artificial neural network includes a further output layer, which is configured to output a degree of an association of the compound to the at least three classes regardless of the productivity and of the frequency of its components. ### Claim 38 (depends on 36) The artificial neural network as recited in claim 36, wherein the artificial neural network includes a further output layer which is configured to output a degree of an association of one of the components to the at least three classes as a function of the productivity and of the frequency of the component. ### Claim 39 A method for classifying term candidates, the method comprising: predefining a feature to an input layer of an artificial neural network; assigning an identifier to the feature from the input layer in a prediction in an output layer of the artificial neural network, the feature characterizing term candidates from a corpus, the corpus including a text from a domain, the identifier characterizing at least three classes for the term candidates that differ from one another, the different classes indicating different degrees of association of the term candidates to the domain. ### Claim 40 (depends on 39) The method as recited in claim 39, wherein a compound and its components for a first feature is predefined to a first input layer of the artificial network, a productivity and a frequency of the components for a second feature is predefined to a second input layer, the output layer is situated downstream from the first and second input layers, and the identifier in the prediction is output as a function of the first feature and of the second feature.