Intent and slot detection for digital assistants
Described herein is a mechanism to adapt a machine learning model used in a language understanding model that has been trained using a first set of user input having a first set of features to effectively operate using user input having a second set of features. Losses are defined based on the first set of features, the second set of features or features common to both the first set and second set. The losses comprise one or more of a source side tagging loss, a reconstruction loss, an adversarial domain classification loss, a non-adversarial domain classification loss, an orthogonality loss, and target side tagging loss. The losses are jointly minimized using a gradient descent method and the resulting coefficients are used to retrain the machine learning model.
1. A method for training a language understanding model, comprising:
accessing a set of initial training data used previously to train a machine learning model of a conversational system and a set of user input data input into the conversational system and received after the initial training data, the set of initial training data representing possible or actual input into the conversational system;
setting the initial set of training data as a source domain and setting the set of user input as a target domain;
creating a representation of a plurality of losses each of a different loss type, each loss based on the source domain, the target domain or both;
defining an aggregate loss comprising the plurality of losses so that the aggregate loss is a function of both the source domain and the target domain;
determine a set of coefficients for the machine learning model using a numerical method, the determined set of coefficients minimizing the aggregate loss within a threshold amount of a local minimum; and
retraining the machine learning model using the determined set of coefficients.
2. The method of claim 1 wherein the plurality of losses comprises one or more of:
a source domain tagging loss;
a reconstruction loss;
an adversarial domain classification loss;
a non-adversarial domain classification loss; and
an orthogonality loss.
3. The method of claim 2 wherein the plurality of losses further comprises the target domain tagging loss.
4. The method of claim 2 wherein the source domain tagging loss is derived based on a first BiLSTM network that induces source domain specific features and a second BiLSTM network that induces domain invariant features.
5. The method of claim 2 wherein the reconstruction loss is derived based on a first BiLSTM network that induces source domain specific features, a second BiLSTM network that induces target domain specific features, and a third BiLSTM network that induces domain invariant features.
6. The method of claim 2 wherein the adversarial domain classification loss is derived based on a first BiLSTM network that induces domain invariant features.
7. The method of claim 2 wherein the non-adversarial domain classification loss is derived based on a first BiLSTM network that induces source domain specific features, and a second BiLSTM network that induces target domain specific features.
8. The method of claim 2 wherein the orthogonality loss is derived based on a first BiLSTM network that induces source domain specific features, a second BiLSTM network that induces target domain specific features, and a third BiLSTM network that induces domain invariant features.
9. The method of claim 1 wherein the initial training data comprises synthetic data created to cold start the machine learning model and wherein the set of user input data comprises data input into the system after the machine learning model has been trained using the synthetic data.
10. The method of claim 1 wherein the initial training data comprises actual data input into the system prior to a time and wherein the set of user input data comprises data input into the system after the time.
11. A system comprising at least a processor and computer executable instructions, that when executed by the at least processor, cause the system to perform operations comprising:
access a set of initial training data used previously to train a machine learning model of a conversational system, the initial training data comprising possible or actual user input into the conversational system;
access a set of user input data input into the conversational system after initial training of the conversational system;
set the initial set of training data as a source domain and set the set of user input as a target domain;
create a representation for each of a plurality of losses each of a different loss type, each loss based on the source domain, the target domain, or both;
define an aggregate loss comprising the plurality of losses;
determine a set of coefficients for the machine learning model using a numerical method, the determined set of coefficients minimizing the aggregate loss within a threshold amount of a local minimum; and
retrain the machine learning model using the determined set of coefficients.
12. The system of claim 11 wherein the plurality of losses comprises one or more of:
a source domain tagging loss;
a reconstruction loss;
an adversarial domain classification loss;
a non-adversarial domain classification loss; and
an orthogonality loss.
13. The system of claim 12 wherein the plurality of losses further comprises a target domain tagging loss.
14. The system of claim 12 wherein the source domain tagging loss is derived based on a first BiLSTM network that induces source domain specific features and a second BiLSTM network that induces domain invariant features.
15. The system of claim 12 wherein the reconstruction loss is derived based on a first BiLSTM network that induces source domain specific features, a second BiLSTM network that induces target domain specific features, and a third BiLSTM network that induces domain invariant features.
16. The system of claim 12 wherein the adversarial domain classification loss is derived based on a first BiLSTM network that induces domain invariant features.
17. The system of claim 12 wherein the non-adversarial domain classification loss is derived based on a first BiLSTM network that induces source domain specific features, and a second BiLSTM network that induces target domain specific features.
18. The system of claim 12 wherein the orthogonality loss is derived based on a first BiLSTM network that induces source domain specific features, a second BiLSTM network that induces target domain specific features, and a third BiLSTM network that induces domain invariant features.
19. A computer storage medium comprising executable instructions that, when executed by a processor of a machine, cause the machine to perform operations comprising:
access a set of initial training data used previously to train a machine learning model of a conversational system, the initial training data comprising possible or actual user input into the conversational system;
access a set of user input data input into the conversational system after initial training of the conversational system;
set the initial set of training data as a source domain and set the set of user input as a target domain;
select two or more loss factors, each of which are a function of the source domain, the target domain, or both, the two or more loss factors comprising:
a source side tagging loss;
a reconstruction loss;
an adversarial domain classification loss;
a non-adversarial domain classification loss; and
an orthogonality loss;
define an aggregate loss comprising the selected loss factors;
determine a set of coefficients for the machine learning model using a numerical method, the determined set of coefficients minimizing the aggregate loss within a threshold amount of a local minimum; and
retrain the machine learning model using the determined set of coefficients.
20. The medium of claim 19 wherein the operations further comprise identify a retraining event and, responsive to the retraining event, execute the operations of claim 19 .