Utterance generation and evaluation
AssigneeAmazon Technologies, Inc.
InventorsVaidyanathan Puthucode Krishnamoorthy, Deepak Babu P R, Ashwin Gopinath, Sethuraman Ramachandran, Ankit Tiwari
Life of this patent
Estimating the term…
Abstract
Devices and techniques are generally described for generating and evaluating utterances. In some examples, an utterance generation and evaluation system can receive intent data and target data. The utterance generation and evaluation system can determine related target names and related intent names and, based on the related target names and related intent names, can generate an utterance phrase. The utterance generation and evaluation system can determine a confidence score associated with the utterance phrase and, based on the confidence score, determine the utterance phrase as a recommended utterance phrase.
Claims18
A method of utterance generation and evaluation comprising:
- receiving item name data associated with an item in an online catalog;
- inputting, into a collaborative filtering model, the item name data;
- receiving, from the collaborative filtering model, a plurality of related item names determined by the collaborative filtering model to be similar to the item name data;
- receiving intent data related to an action associated with the item;
- inputting, into the collaborative filtering model, the intent data;
- receiving, from the collaborative filtering model, a plurality of related user intents determined by the collaborative filtering model to be similar to the intent data;
- determining an utterance phrase by combining an individual related item name and an individual related user intent;
- inputting, into an utterance audio model, the utterance phrase;
- receiving, from the utterance audio model, a plurality of utterance audio samples;
- inputting, into a machine learned model, the plurality of utterance audio samples;
- receiving, from the machine learned model, a plurality of confidence scores, wherein an individual confidence score of the plurality of confidence scores is associated with an individual utterance audio sample of the plurality of utterance audio samples;
- determining that the individual confidence score meets or exceeds a confidence score threshold;
- and
- determining the utterance phrase as a candidate utterance phrase.
A method comprising:
- receiving target data describing content included in a database;
- inputting the target data into a first machine learning model executed by at least one processor;
- determining, by the first machine learning model based at least in part on the target data, one or more related target names;
- receiving intent data related to an action associated with the content;
- inputting the intent data into the first machine learning model executed by the at least one processor;
- determining, by the first machine learning model based at least in part on the intent data, one or more related intents;
- generating, based at least in part on the one or more related target names and the one or more related intents, utterance data representing an utterance phrase;
- inputting the utterance data into a second machine learned model executed by the at least one processor;
- determining, by the second machine learned model based at least in part on the utterance data, a confidence score associated with the utterance data;
- and
- determining, based at least in part on the confidence score, that the utterance phrase is a recommended utterance phrase;
- and
- sending, to a remote computing device, the utterance data.
A non transitory computer readable medium storing instructions executable by a processor, wherein the instructions, when executed, cause the processor to perform operations:
- receiving target data describing content included in a database;
- inputting the target data into a first machine learning model executed by at least a first processor;
- determining, by the first machine learning model based at least in part on the target data, one or more related target names;
- receiving intent data related to an action associated with the content;
- inputting the intent data into the first machine learning model executed by at least the first processor;
- determining, by the first machine learning model based at least in part on the intent data, one or more related intents;
- generating, based at least in part on the one or more related target names and the one or more related intents, utterance data representing an utterance phrase;
- inputting the utterance data into a second machine learned model;
- determining, by the second machine learned model based at least in part on the utterance data, a confidence score associated with the utterance data;
- and
- determining, based at least in part on the confidence score, that the utterance phrase is a recommended utterance phrase;
- and
- sending, to a remote computing device, the utterance data.
Dependencies are inferred from the claim text.