Back to results
US 11,386,890 B1Granted
Natural language understanding
AssigneeAmazon Technologies, Inc.
InventorsXing Fan, Zheng Chen, Yuan Ling, Lambert Leo Mathias, Chenlei Guo
Life of this patent
Estimating the term…
Abstract
A system is provided for reducing friction during user interactions with a natural language processing system, such as voice assistant systems. The system determines a pre-trained model using dialog session data corresponding to multiple user profiles. The system determines a fine-tuned model using the pre-trained model and a fine-tuning dataset that corresponds to a particular task, such as query rewriting. The system uses the fine-tuned model to process a user input and determine an alternative representation of the input that can result in a desired response from the natural language processing system.
Claims20
A computer-implemented method comprising:
- during a first time period:
- determining first dialog session data corresponding to a first user profile, the first dialog session data representing at least a first utterance, a second utterance, and a system-generated response;
- determining second dialog session data corresponding to a second user profile;
- determining a first dataset including at least the first dialog session data and the second dialog session data;
- processing the first dataset to determine first model data corresponding to a first model, the first model configured to predict a subsequent user input for a first user input in a dialog session;
- determining first pair data representing a third utterance that results in an undesired response and first text data corresponding to an alternative representation of the third utterance, the first text data resulting in a desired response;
- determining a second pair data representing a fifth utterance that results in an undesired response and second text data corresponding to an alternative representation of the fifth utterance, the second text data resulting in a desired response;
- determining a second dataset including the first pair data and the second pair data;
- and
- processing the first model data and the second dataset to determine second model data corresponding to a second model, the second model configured to predict an alternative user input for a second user input that results in an undesired response;
- during a second time period after the first time period:
- receiving audio data representing an input utterance;
- processing the audio data using automatic speech recognition (ASR) processing to determine first input text data representing the input utterance;
- processing, using the second model, the first input text data to determine output model data;
- determining, using the output model data, second input text data representing an alternative representation for the input utterance, the alternative representation known to result in a desired response;
- and
- generating, using the second input text data, output data responsive to the input utterance.
14 more elements
A computer-implemented method comprising:
- receiving first input data;
- processing the first input data using a first model to determine first data representing a first alternative representation of the first input data, the first alternative representation known to result in a desired response, wherein the first model is determined by:
- determining first dialog data corresponding to multiple profiles;
- determining, using the first dialog data, a first past user input;
- determining, using the first dialog data, a second past user input;
- determining that the second past user input is an alternative representation of the first past user input;
- determining a first dataset including at least an association between the first past user input and the second past user input;
- and
- determining the first model using the first dataset;
- and
- generating, using the first data, first output data responsive to the first input data.
8 more elements
A system comprising:
- at least one processor;
- and
- at least one memory including instructions that, when executed by the at least one processor, cause the system to:
- receive first input data;
- process the first input data using a first model to determine first data representing a first alternative representation of the first input data, the first alternative representation known to result in a desired response;
- and
- generate, using the first data, first output data responsive to the first input data, and
- wherein the system determines the first model by executing the instructions that cause the system to:
- determine first dialog data corresponding to multiple profiles;
- determine, using the first dialog data, a first past user input;
- determine, using the first dialog data, a second past user input;
- determine that the second past user input is an alternative representation of the first past user input;
- determine a first dataset including at least an association between the first past user input and the second past user input;
- and
- determine the first model using the first dataset.
12 more elements
Dependencies are inferred from the claim text.