US 12,592,226 B1Granted

Natural language processing skill candidate determination

AssigneeAmazon Technologies, Inc.

InventorHassen Dhrif

Life of this patent

Estimating the term…

Loading patent…

Abstract

Devices and techniques are generally described for natural language processing interfaces. In various examples, first natural language data may be received from an input device. First embedding data representing the first natural language data may be generated. A nearest neighbor algorithm may determine first data representing similarity between the first embedding data and second embedding data, the second embedding data representing second natural language data associated with a first skill. The nearest neighbor algorithm may determine second data representing similarity between the first embedding data and third embedding data, the third embedding data representing third natural language data associated with a second skill. First output data that indicates that the first skill and the second skill are candidates for processing the first natural language data may be generated.

Claims18

Dependent claims

A computer-implemented method comprising:

receiving, from an input device, a first natural language input;
generating, using a pre-trained transformer model, sentence embedding data representing the first natural language input;
generating, by searching a positive search index using a nearest neighbor algorithm, a first similarity score representing similarity between the sentence embedding data and first embedding data stored in the positive search index, the first embedding data representing a first utterance to which a first application is able to respond;
generating, by searching the positive search index using the nearest neighbor algorithm, a second similarity score representing similarity between the sentence embedding data and second embedding data stored in the positive search index, the second embedding data representing a second utterance to which a second application is able to respond;
generating, by searching a negative search index using the nearest neighbor algorithm, a third similarity score representing similarity between the sentence embedding data and third embedding data stored in the negative search index, wherein the negative search index comprises embedding data associated with respective utterances that have been specified as utterances to which the first application is unable to respond;
determining, using the first similarity score and the second similarity score, that the sentence embedding data is more similar to the first embedding data than the second embedding data;
determining, using the third similarity score and the first similarity score, that the sentence embedding data is more similar to the third embedding data than the first embedding data;
determining a set of candidate applications for processing the first natural language input, wherein the first application is excluded from the set of candidate applications based on the sentence embedding data being more similar to the third embedding data than the first embedding data;
generating first data indicating that the second application is a first candidate of a set of candidate applications for processing the first natural language input;
determining the second application for processing the first natural language input from among the set of candidate applications based on the first data;
and
sending data representing the first natural language input to the second application, wherein the second application is effective to perform a first action in response to the first natural language input.

9 more elements

A method comprising:

receiving, from an input device, first natural language data;
generating first embedding data representing the first natural language data;
generating, based at least in part on searching a first search index using a nearest neighbor algorithm, first data representing similarity between the first embedding data and second embedding data stored in the first search index, the second embedding data representing second natural language data to which a first application is able to respond;
generating, based at least in part on searching the first search index using the nearest neighbor algorithm, second data representing similarity between the first embedding data and third embedding data stored in the first search index, the third embedding data representing third natural language data to which a second application is able to respond;
generating, based at least in part on searching a second search index using the nearest neighbor algorithm, third data representing similarity between the first embedding data and fourth embedding data stored in the second search index, wherein the second search index is associated with natural language inputs to which the first application is unable to respond;
generating first output data indicating that the first application and the second application are candidates for processing the first natural language data;
sending request data representing the first natural language data to the first application based at least in part on the first data and the third data, wherein the first application is effective to determine an action based at least in part on the request data;
and
receiving, from the first application, first response data generated in response to the first application receiving the request data.

6 more elements

A system comprising:

at least one processor;
and
non-transitory computer-readable memory storing instructions that, when executed by the at least one processor, are effective to:
receive, from an input device, first natural language data;
generate first embedding data representing the first natural language data;
generate, based at least in part on searching a first search index using a nearest neighbor algorithm, first data representing similarity between the first embedding data and second embedding data stored in the first search index, the second embedding data representing second natural language data to which a first application is able to respond;
generate, based at least in part on searching the first search index using the nearest neighbor algorithm, second data representing similarity between the first embedding data and third embedding data stored in the first search index, the third embedding data representing third natural language data to which a second application is able to respond;
generate, based at least in part on searching a second search index using the nearest neighbor algorithm, third data representing similarity between the first embedding data and fourth embedding data stored in the second search index, wherein the second search index comprises embeddings representing natural language inputs to which the first application is unable to respond;
generate first output data indicating that the first application and the second application are candidates for processing the first natural language data;
send request data representing the first natural language data to the first application, based on the first data and the third data, wherein the first application is effective to determine an action based at least in part on the request data representing the first natural language data;
and
receive, from the first application, first response data generated in response to the first application receiving the request data.

9 more elements

Dependencies are inferred from the claim text.