Systems and methods relating to knowledge distillation in natural language processing models

1. A method for creating a student model from a teacher model for use in knowledge distillation, the method comprising:

providing the teacher model, wherein:

the teacher model comprises a neural network having a plurality of layers; and

the teacher model is trained on a first training dataset;

generating candidate student models, each of the candidate student models comprising a model having a unique permutation of layers derived by randomly selecting one or more layers of the plurality of layers of the teacher model for removing;

generating a second training dataset, the second training dataset comprising a randomly selected data from the first training dataset;

for each of the candidate student models:

providing the second training dataset as inputs to the candidate student model;

recording outputs generated by the candidate student model from the second training data set; and

based on the recorded outputs, evaluating a performance of the candidate student model according to a predetermined model evaluation criterion;

determining which of the candidate student models performed best among the candidate student models based on the predetermined model evaluation criterion; and

identifying a preferred candidate student model as being the candidate student model that performed best.

2. The method of claim 1 , wherein the predetermined model evaluation criterion comprises an output accuracy wherein the output of the candidate student model is compared to an output of the teacher model.

3. The method of claim 1 , wherein the predetermined model evaluation criterion comprises a Bayesian Information Criterion.

4. The method of claim 1 , further comprising the step of:

training the preferred candidate student model so that an output of the preferred candidate student model over a transfer dataset mimics an output of the teacher model over the transfer dataset to within an acceptable loss of validity as defined by a predetermined loss function.

5. The method of claim 4 , further comprising the step of:

employing the preferred candidate student model for a natural language processing task in a contact center, the natural language processing task comprising at least one of natural language inference, sentiment classification, and semantic textual similarity.

6. The method of claim 5 , wherein the teacher model comprises a bidirectional encoder representations from transforms architecture.

7. A system for creating a student model from a teacher model for use in knowledge distillation, the system comprising:

at least one processor; and

at least one memory, wherein the at least one memory stores instructions that, when executed by the at least one processor, causes the at least one processor to:

provide the teacher model, wherein:

the teacher model comprises a neural network having a plurality of layers; and

the teacher model is trained on a first training dataset;

generate candidate student models, each of the candidate student models comprising a model having a unique permutation of layers derived from random selection of one or more layers of the plurality of layers of the teacher model for removal;

generate a second training dataset, the second training dataset comprising a randomly selected data from the first training dataset;

for each of the candidate student models:

provide the second training dataset as inputs to the candidate student model;

record outputs generated by the candidate student model from the second training data set; and

based on the recorded outputs, evaluate a performance of the candidate student model according to a predetermined model evaluation criterion;

determine which of the candidate student models performed best among the candidate student models based on the predetermined model evaluation criterion; and

identify a preferred candidate model as being the candidate student model that performed best.

8. The system of claim 7 , wherein the predetermined model evaluation criterion comprises an output accuracy in which the output of the candidate student model is compared to an output of the teacher model.

9. The system of claim 7 , wherein the predetermined model evaluation criterion comprises a Bayesian Information Criterion.

10. The system of claim 7 , wherein the memory further stores instructions that, when executed by the processor, cause the processor to perform the step of:

training the preferred candidate student model so that an output of the preferred candidate student model over a transfer dataset mimics an output of the teacher model over the transfer dataset to within an acceptable loss of validity as defined by a predetermined loss function.

11. The system of claim 10 , wherein the memory further stores instructions that, when executed by the processor, cause the processor to perform the step of:

employing the preferred candidate student model for a natural language processing task in a contact center, the natural language processing task comprising at least one of natural language inference, sentiment classification, and semantic textual similarity.

12. The system of claim 11 , wherein the first model comprises a bidirectional encoder representations from transforms architecture.

13. At least one non-transitory computer-readable storage medium having computer instructions stored therein, which when executed by a computing system, causes the computing system to:

provide a teacher model, wherein:

the teacher model comprises a neural network having a plurality of layers; and

the teacher model is trained on a first training dataset;

generate candidate student models, each of the candidate student models comprising a model having a unique permutation of layers derived from random selection of one or more layers of the plurality of layers of the teacher model for removal;

generate a second training dataset, the second training dataset comprising a randomly selected data from the first training dataset;

for each of the candidate student models:

provide the second training dataset as inputs to the candidate student model;

record outputs generated by the candidate student model from the second training data set; and

based on the recorded outputs, evaluate a performance of the candidate student model according to a predetermined model evaluation criterion;

determine which of the candidate student models performed best among the candidate student models based on the predetermined model evaluation criterion; and

identify a preferred candidate model as being the candidate student model that performed best.

14. The at least one non-transitory computer-readable storage medium of claim 13 , wherein the predetermined model evaluation criterion comprises an output accuracy wherein the output of the candidate student model is compared to an output of the teacher model.

15. The at least one non-transitory computer-readable storage medium of claim 13 , wherein the predetermined model evaluation criterion comprises a Bayesian Information Criterion.

16. The at least one non-transitory computer-readable storage medium of claim 13 , wherein the computer instructions stored therein, when executed by the computing system, further causes the computing system computer to:

train the preferred candidate student model so that an output of the preferred candidate student model over a transfer dataset mimics an output of the teacher model over the transfer dataset to within an acceptable loss of validity as defined by a predetermined loss function.

17. The at least one non-transitory computer-readable storage medium of claim 16 , wherein the computer instructions stored therein, when executed by the computing system, further causes the computing system to:

employ the preferred candidate student model for a natural language processing task in a contact center, the natural language processing task comprising at least one of natural language inference, sentiment classification, and semantic textual similarity.

18. The at least one non-transitory computer-readable storage medium of claim 17 , wherein the first model comprises a bidirectional encoder representations from transforms architecture.