Evaluating natural language processing components
Techniques for evaluating a natural language understanding (NLU) component and determining an action to resolve an issue processing a user input are described. The system determines which component is invoked by a baseline NLU component is processing the user input, and which component is invoked by an updated NLU component. Based on that information, the system selects the action to resolve the updated NLU component generating an undesired response to the user input.
1. A computer-implemented method, comprising:
receiving first data corresponding to processing of a first user input using a first component, the first data including first output data;
receiving second data corresponding to processing of the first user input using a first updated component, the second data including second output data different from the first output data;
determining the second data results in an undesired response; and
determining a second updated component using the first updated component, the determining of the second updated component being based at least in part on processing of the first data with respect to the second data.
2. The computer-implemented method of claim 1 , further comprising:
determining the second output data corresponds to a message indicating an error in processing the first user input.
3. The computer-implemented method of claim 1 , further comprising:
determining that processing of the first user input by the second updated component results in a desired response; and
storing the second updated component for processing of future user inputs.
4. The computer-implemented method of claim 1 , wherein:
the first component comprises a first model and a second model; and
the first updated component comprises the first model and a third model, the third model corresponding to an updated version of the second model.
5. The computer-implemented method of claim 4 , further comprising:
determining the undesired response corresponds to processing by the third model.
6. The computer-implemented method of claim 5 , further comprising:
determining that first grammar data corresponding to the second model is different than second grammar data corresponding to the third model; and
determining third grammar data by updating the second grammar data.
7. The computer-implemented method of claim 1 , wherein:
the first component comprises a first model and a second model;
the first updated component comprises a third model and a fourth model; and
the second updated component comprises the third model and a fifth model, the fifth model corresponding to an updated version of the fourth model.
8. The computer-implemented method of claim 1 , wherein;
the first component comprises a first model and a second model;
the first updated component comprises a third model; and
the method further comprises:
determining that the first data indicates that the first model was invoked during processing of the first user input,
determining that the second data indicates that the third model was invoked during processing of the first user input,
processing the first output data with respect to the second output data,
based at least in part on the processing of the first output data with respect to the second output data, determining training data to include a set of user inputs corresponding to the first user input, the set of user inputs associated with a positive label, and
training, using the training data, the first updated component to determine the second updated component.
9. The computer-implemented method of claim 8 , further comprising:
receiving third data representing processing of the first user input using the second updated component, the second updated component comprising a third model and a fourth model, the third data comprising third output data;
determining that processing of the first user input by the second updated component results in the undesired response;
determining that the third data indicates that the third model was invoked during processing of the first user input;
processing the first output data with respect to the third output data; and
generating a fifth model corresponding to the first user input, the fifth model to be included in the second updated component.
10. The computer-implemented method of claim 1 , wherein:
the first component comprises a first model and a second model;
the first updated component comprises a third model; and
the method further comprises:
determining that the first data indicates that the first model was invoked during processing of the first user input,
determining that the second data indicates that the third model was invoked during processing of the first user input,
processing the first output data with respect to the second output data,
based at least in part on the processing of the first output data with respect to the second output data, determining training data to include a set of user inputs corresponding to the first user input, the set of user inputs associated with a negative label, and
training, using the training data, the first updated component to determine the second updated component.
11. A system comprising:
at least one processor; and
at least one memory comprising instructions that, when executed by the at least one processor, cause the system to:
receive first data corresponding to processing of a first user input using a first component, the first data including first output data;
receive second data corresponding to processing of the first user input using a first updated component, the second data including second output data different from the first output data;
determine the second data results in an undesired response; and
determine a second updated component using the first updated component, wherein determination of the second updated component being based at least in part on processing of the first data with respect to the second data.
12. The system of claim 11 , wherein the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
determine the second output data corresponds to a message indicating an error in processing the first user input.
13. The system of claim 11 , wherein the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
determine that processing of the first user input by the second updated component results in a desired response; and
store the second updated component for processing of future user inputs.
14. The system of claim 11 , wherein:
the first component comprises a first model and a second model; and
the first updated component comprises the first model and a third model, the third model corresponding to an updated version of the second model.
15. The system of claim 14 , wherein the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
determine the undesired response corresponds to processing by the third model.
16. The system of claim 15 , wherein the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
determine that first grammar data corresponding to the second model is different than second grammar data corresponding to the third model; and
determine third grammar data by updating the second grammar data.
17. The system of claim 11 , wherein:
the first component comprises a first model and a second model;
the first updated component comprises a third model and a fourth model; and
the second updated component comprises the third model and a fifth model, the fifth model corresponding to an updated version of the fourth model.
18. The system of claim 11 , wherein:
the first component comprises a first model and a second model;
the first updated component comprises a third model; and
the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
determine that the first data indicates that the first model was invoked during processing of the first user input,
determine that the second data indicates that the third model was invoked during processing of the first user input,
process the first output data with respect to the second output data,
based at least in part on the processing of the first output data with respect to the second output data, determine training data to include a set of user inputs corresponding to the first user input, the set of user inputs associated with a positive label, and
train, using the training data, the first updated component to determine the second updated component.
19. The system of claim 18 , wherein the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
receive third data representing processing of the first user input using the second updated component, the second updated component comprising a third model and a fourth model, the third data comprising third output data;
determine that processing of the first user input by the second updated component results in the undesired response;
determine that the third data indicates that the third model was invoked during processing of the first user input;
process the first output data with respect to the third output data; and
generate a fifth model corresponding to the first user input, the fifth model to be included in the second updated component.
20. The system of claim 11 , wherein:
the first component comprises a first model and a second model;
the first updated component comprises a third model; and
the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
determine that the first data indicates that the first model was invoked during processing of the first user input,
determine that the second data indicates that the third model was invoked during processing of the first user input,
process the first output data with respect to the second output data,
based at least in part on the processing of the first output data with respect to the second output data, determine training data to include a set of user inputs corresponding to the first user input, the set of user inputs associated with a negative label, and
train, using the training data, the first updated component to determine the second updated component.