System and method for language processing using adaptive regularization
A system and method incorporate prior knowledge into the optimization and regularization of a classification and regression model. The optimization may be a regularization process and the prior knowledge may be incorporated through adjustment of a cost function. A method of at least one processor developing a classification and regression model may be provided. The method may be implemented by at least one processor that implements classification and regression model functionality, including receiving training data and adjusting the model according to the training data; testing the classification and regression model; and employing prior knowledge during an optimization of the classification and regression model. The regularizing can include adjusting feature weights according to prior knowledge. In various embodiments, such systems and methods can be used in the processing of language inputs, e.g., speech and/or text inputs, to achieve greater interpretation accuracy.
1. A language processing method, carried out by at least one processor having access to at least one computer storage device, the method comprising:
forming or accessing a classification and regression model, wherein the classification and regression model is a natural language understanding model;
receiving training data;
adjusting the classification and regression model according to the training data; and
employing prior knowledge to optimize the classification and regression model, including applying feature weights to one or more features of the classification and regression model, to form an optimized classification and regression model, wherein the feature weights include natural language understanding feature weights, and wherein the optimization includes the at least one processor regularizing the classification and regression model using different regularization values according to feature types of the one or more features of the classification and regression model, and wherein the regularizing includes adjusting a regularizing cost function to incorporate at least some of the prior knowledge to give higher priority to a feature of the one or more features that is expected to be more general within training data.
2. The method of claim 1 , wherein the regularizing includes the at least one processor adjusting the feature weights according to at least some of the prior knowledge.
3. The method of claim 1 , wherein the feature weights further include machine learning feature weights.
4. The method of claim 1 , wherein the regularizing includes the at least one processor adjusting a cost function to incorporate at least some of the prior knowledge.
5. The method of claim 4 , further comprising the at least one processor adjusting the cost function to give higher priority to a feature that is expected to be more general within training data.
6. A language processing system, comprising:
at least one processor configured to form or access a classification and regression model, wherein the classification and regression model is a natural language understanding model;
the at least one processor further configured to receive training data;
the at least one processor further configured to adjust the classification and regression model according to the training data; and
the at least one processor further configured to employ prior knowledge and to apply feature weights to one or more features of the classification and regression model to form an optimized classification and regression model, wherein the feature weights include natural language understanding feature weights, and wherein the optimization includes the at least one processor regularizing the classification and regression model using different regularization values according to feature types of the one or more features of the classification and regression model, and wherein the regularizing includes adjusting a regularizing cost function to incorporate at least some of the prior knowledge to give higher priority to a feature of the one or more features that is expected to be more general within training data.
7. The system of claim 6 , wherein the at least one processor is configured to adjust feature weights according to at least some of the prior knowledge during the regularization.
8. The system of claim 6 , wherein the at least one processor is configured to employ machine learning feature weights when adjusting the feature weights.
9. The system of claim 6 , wherein the at least one processor is configured to adjust a cost function to incorporate at least some of the prior knowledge during regularization.
10. The system of claim 9 , wherein the at least one processor is configured to adjust the cost function to give higher priority to a feature that is expected to be more general within training data.
11. A method employing natural language understanding, carried out by at least one processor having access to at least one computer storage device, the method comprising:
forming or accessing a classification and regression model, wherein the classification and regression model is a natural language understanding model;
receiving training data;
adjusting the classification and regression model according to the training data;
testing the classification and regression model;
optimizing, employing prior knowledge, the classification and regression model, including applying feature weights to one or more features of the classification and regression model, to form an optimized classification and regression model, wherein the feature weights include natural language understanding feature weights; and
receiving operational data and employing the optimized classification and regression model to classify the operational data to adapt the feature weights, and wherein the optimization includes the at least one processor regularizing the classification and regression model using different regularization values according to feature types of the one or more features of the classification and regression model, and wherein the regularizing includes adjusting a regularizing cost function to incorporate at least some of the prior knowledge to give higher priority to a feature of the one or more features that is expected to be more general within training data.
12. The method of claim 11 , wherein the classification of the operational data is related to the intent of the operational data.
13. The method of claim 11 , wherein the regularizing includes the at least one processor adjusting the feature weights according to at least some of the prior knowledge.
14. The method of claim 11 , wherein the regularizing includes adjusting a cost function to incorporate at least some of the prior knowledge and/or to give higher priority to a feature that is expected to be more general within training data.