US 11,144,718 B2Granted

Adaptable processing components

AssigneeInternational Business Machines Corporation

InventorsYoungja Park, Siddharth A. Patwardhan

Life of this patent

Estimating the term…

Abstract

In configuring a processing system with an application made up of machine learning components, where the application has been trained on a set of training data, the application is executed on the processing system using another set of training data. Outputs of the application produced from the other set of training data identified that concur with ground truth data are identified. The components are adapted to produce outputs of the application that concur with the ground truth data using the identified outputs of the application.

Claims19

Dependent claims

A method of configuring a processing system with an application including a natural language processing core utilizing a plurality of machine learning components, the method comprising:

training each machine learning component of the plurality of machine learning components individually with respective training data for subject matter of a first domain to perform a corresponding natural language processing task;
training the application on an initial set of training data for the subject matter of the first domain;
executing the application with the trained machine learning components on the processing system using another set of training data for different subject matter of a second different domain to produce a plurality of different outputs from the application for each of a plurality of elements of the other set of training data, wherein each different output from the application is based on a corresponding one of a first plurality of candidate outputs generated for each of the elements by a trained machine learning component of the plurality of trained machine learning components performing the corresponding natural language processing task;
identifying, for each of the elements of the other set of training data, one or more outputs of the application among the different outputs of the application that concur with ground truth data;
identifying the candidate outputs of the trained machine learning component that produce the identified outputs concurring with the ground truth data;
weighting each of the first plurality of candidate outputs of the trained machine learning component, wherein weighting for the identified candidate outputs is greater relative to weighting for other candidate outputs, and the weighting each of the first plurality of candidate outputs comprises:
training a component model for the trained machine learning component to produce weight vectors for the first plurality of candidate outputs and updating the component model to adjust the weight vectors based on the identified outputs of the application;
and
applying corresponding weight vectors of the component model to feature vectors of the first plurality of candidate outputs to produce a score for each of the first plurality of candidate outputs;
and
adapting the natural language processing core to the second different domain by adapting the plurality of trained machine learning components to select based on the weighting score, a candidate output from the trained machine learning component that provides an identified output concurring with the ground truth data and to propagate the selected candidate output through remaining machine learning components to produce a result of the application for an inquiry of the different subject matter of the different second domain.

8 more elements

An apparatus to configure a processing system with an application including a natural language processing core utilizing a plurality of machine learning components, the apparatus comprising:

a processor configured to:
train each machine learning component of the plurality of machine learning components individually with respective training data for subject matter of a first domain to perform a corresponding natural language processing task;
train the application on an initial set of training data for the subject matter of the first domain;
execute the application with the trained machine learning components on the processing system using another set of training data for different subject matter of a second different domain to produce a plurality of different outputs from the application for each of a plurality of elements of the other set of training data, wherein each different output from the application is based on a corresponding one of a first plurality of candidate outputs generated for each of the elements by a trained machine learning component of the plurality of trained machine learning components performing the corresponding natural language processing task;
identify, for each of the elements of the other set of training data, one or more outputs of the application among the different outputs of the application that concur with ground truth data;
identify the candidate outputs of the trained machine learning component that produce the identified outputs concurring with the ground truth data;
weight each of the first plurality of candidate outputs of the trained machine learning component, wherein weighting for the identified candidate outputs is greater relative to weighting for other candidate outputs, and the weighting each of the first plurality of candidate outputs comprises:
training a component model for the trained machine learning component to produce weight vectors for the first plurality of candidate outputs and updating the component model to adjust the weight vectors based on the identified outputs of the application;
and
applying corresponding weight vectors of the component model to feature vectors of the first plurality of candidate outputs to produce a score for each of the first plurality of candidate outputs;
and
adapt the natural language processing core to the second different domain by adapting the plurality of trained machine learning components to select based on the score, a candidate output from the trained machine learning component that provides an identified output concurring with the ground truth data and to propagate the selected candidate output through remaining machine learning components to produce a result of the application for an inquiry of the different subject matter of the different second domain.

9 more elements

A computer program product for configuring a processing system with an application including a natural language processing core utilizing a plurality of machine learning components, the computer program product comprising one or more computer readable storage media collectively having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:

train each machine learning component of the plurality of machine learning components individually with respective training data for subject matter of a first domain to perform a corresponding natural language processing task;
train the application on an initial set of training data for the subject matter of the first domain;
execute the application with the trained machine learning components on the processing system using another set of training data for different subject matter of a second different domain to produce a plurality of different outputs from the application for each of a plurality of elements of the other set of training data, wherein each different output from the application is based on a corresponding one of a first plurality of candidate outputs generated for each of the elements by a trained machine learning component of the plurality of trained machine learning components performing the corresponding natural language processing task;
identify, for each of the elements of the other set of training data, one or more outputs of the application among the different outputs of the application that concur with ground truth data;
identify the candidate outputs of the trained machine learning component that produce the identified outputs concurring with the ground truth data;
weight each of the first plurality of candidate outputs of the trained machine learning component, wherein weighting for the identified candidate outputs is greater relative to weighting for other candidate outputs, and the weighting each of the first plurality of candidate outputs comprises:
training a component model for the trained machine learning component to produce weight vectors for the first plurality of candidate outputs and updating the component model to adjust the weight vectors based on the identified outputs of the application;
and
applying corresponding weight vectors of the component model to feature vectors of the first plurality of candidate outputs to produce a score for each of the first plurality of candidate outputs;
and
adapt the natural language processing core to the second different domain by adapting the plurality of trained machine learning components to select based on the score, a candidate output from the trained machine learning component that provides an identified output concurring with the ground truth data and to propagate the selected candidate output through remaining machine learning components to produce a result of the application for an inquiry of the different subject matter of the different second domain.

8 more elements

Dependencies are inferred from the claim text.

patentgrep

Loading patent…

Abstract

A method of configuring a processing system with an application including a natural language processing core utilizing a plurality of machine learning components, the method comprising:

training each machine learning component of the plurality of machine learning components individually with respective training data for subject matter of a first domain to perform a corresponding natural language processing task;
training the application on an initial set of training data for the subject matter of the first domain;
executing the application with the trained machine learning components on the processing system using another set of training data for different subject matter of a second different domain to produce a plurality of different outputs from the application for each of a plurality of elements of the other set of training data, wherein each different output from the application is based on a corresponding one of a first plurality of candidate outputs generated for each of the elements by a trained machine learning component of the plurality of trained machine learning components performing the corresponding natural language processing task;
identifying, for each of the elements of the other set of training data, one or more outputs of the application among the different outputs of the application that concur with ground truth data;
identifying the candidate outputs of the trained machine learning component that produce the identified outputs concurring with the ground truth data;
weighting each of the first plurality of candidate outputs of the trained machine learning component, wherein weighting for the identified candidate outputs is greater relative to weighting for other candidate outputs, and the weighting each of the first plurality of candidate outputs comprises:
training a component model for the trained machine learning component to produce weight vectors for the first plurality of candidate outputs and updating the component model to adjust the weight vectors based on the identified outputs of the application;
and
applying corresponding weight vectors of the component model to feature vectors of the first plurality of candidate outputs to produce a score for each of the first plurality of candidate outputs;
and
adapting the natural language processing core to the second different domain by adapting the plurality of trained machine learning components to select based on the weighting score, a candidate output from the trained machine learning component that provides an identified output concurring with the ground truth data and to propagate the selected candidate output through remaining machine learning components to produce a result of the application for an inquiry of the different subject matter of the different second domain.

8 more elements