Apparatus and method for model adaptation for spoken language understanding
An apparatus and a method are provided for building a spoken language understanding model. Labeled data may be obtained for a target application. A new classification model may be formed for use with the target application by using the labeled data for adaptation of an existing classification model. In some implementations, the existing classification model may be used to determine the most informative examples to label.
1. A method comprising:
receiving a first language model for a first domain having a first training set, wherein the first language model is used in a spoken dialog application;
identifying a second language model for a second domain, wherein the second language model is smaller than the first language model;
selectively sampling training data from the second language model, to yield selectively sampled training data;
concatenating, via a processor, the first language model by replacing a portion of the first training set in the first language model with the selectively sampled training data, to yield a concatenated first language model;
replacing the first language model with the concatenated first language model for use in the spoken dialog application; and
receiving audible speech from a user via the spoken dialog application; and
converting, via the spoken dialog application, the audible speech into text via the concatenated first language model.
2. The method of claim 1 , wherein concatenating the first language model further comprises:
determining a distance from the first language model to the second language model; and
modifying the first language model using the distance.
3. The method of claim 2 , wherein the first language model is a speech processing model and wherein the distance comprises a logistic loss function.
4. The method of claim 1 , further comprising:
obtaining confidence scores from the first language model and a modified first language model; and
engaging in active learning using the confidence scores.
5. The method of claim 1 , wherein modifying the first language model utilizes one of a Boosting algorithm, a Naïve Bayes classifier, a linear model interpolation, and a Bayesian adaptation.
6. The method of claim 5 , wherein probability distributions correspond to an existing language model probability distribution and a modified language model probability distribution.
7. The method of claim 1 , further comprising labeling future utterances using a modified first language model.
8. A system comprising:
a processor; and
a computer-readable storage medium having instructions stored which, when executed by the processor, cause the processor to perform operations comprising:
receiving a first language model for a first domain having a first training set, wherein the first language model is used in a spoken dialog application;
identifying a second language model for a second domain, wherein the second language model is smaller than the first language model;
selectively sampling training data associated with the second language model, to yield selectively sampled training data;
concatenating, via a processor, the first language model by replacing a portion of the first training set in the first language model with the selectively sampled training data, to yield a concatenated first language model;
replacing the first language model with the concatenated first language model for use in the spoken dialog application; and
receiving audible speech from a user via the spoken dialog application; and
converting, via the spoken dialog application, the audible speech into text via the concatenated first language model.
9. The system of claim 8 , wherein concatenating the first language model further comprises:
determining a distance from the first language model to the second language model; and
modifying the first language model using the distance.
10. The system of claim 9 , wherein the first language model is a speech processing model and wherein the distance comprises a logistic loss function.
11. The system of claim 8 , the computer-readable storage medium having additional instructions stored which, when executed by the processor, cause the processor to perform operations comprising:
obtaining confidence scores from the first language model and a modified first language model; and
engaging in active learning using the confidence scores.
12. The system of claim 8 , wherein modifying the first language model utilizes one of a Boosting algorithm, a Naïve Bayes classifier, a linear model interpolation, and a Bayesian adaptation.
13. The system of claim 12 , wherein probability distributions correspond to an existing language model probability distribution and a modified language model probability distribution.
14. The system of claim 8 , the computer-readable storage medium having additional instructions stored which, when executed by the processor, cause the processor to perform operations comprising labeling future utterances using a modified first language model.
15. A device having instructions stored which, when executed by a computing device, cause the computing device to perform operations comprising:
receiving a first language model for a first domain having a first training set, wherein the first language model is used in a spoken dialog application;
identifying a second language model for a second domain, wherein the second language model is smaller than the first language model;
selectively sampling training data associated with the second language model, to yield selectively sampled training data;
concatenating the first language model by replacing a portion of the first training set in the first language model with the selectively sampled training data, to yield a concatenated first language model;
replacing the first language model with the concatenated first language model for use in the spoken dialog application; and
receiving audible speech from a user via the spoken dialog application; and
converting, via the spoken dialog application, the audible speech into text via the concatenated first language model.
16. The device of claim 15 , wherein concatenating the first language model further comprises:
determining a distance from the first language model to the second language model; and
modifying the first language model using the distance.
17. The device of claim 16 , wherein the first language model is a speech processing model and wherein the distance comprises a logistic loss function.