Speech recognition semantic classification training

• US 20100023331A1
• Filed: 07/15/2009
• Published: 01/28/2010
• Est. Priority Date: 07/17/2008
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Peters, Douglas, Duta, Nicolae, Tremblay, Rèal, Mauro, Andy
• Granted Patent: US 8,781,833 B2
• Time in Patent Office: Days
• US Class Current: 704/257
• CPC Class Codes: G10L 15/01 Assessment or evaluation of speech recognition systems
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 2015/0636 Threshold criteria for the updating

Abstract

An automated method is described for developing an automated speech input semantic classification system such as a call routing system. A set of semantic classifications is defined for classification of input speech utterances, where each semantic classification represents a specific semantic classification of the speech input. The semantic classification system is trained from training data having little or no in-domain manually transcribed training data, and then operated to assign input speech utterances to the defined semantic classifications. Adaptation training data based on input speech utterances is collected with manually assigned semantic labels. When the adaptation training data satisfies a pre-determined adaptation criteria, the semantic classification system is automatically retrained based on the adaptation training data.

Claims

1. A method for developing an automated speech input semantic classification system, the method comprising:

defining a set of semantic classifications for classification of input speech utterances, each semantic classification representing a specific semantic classification of the speech input;

training the semantic classification system from training data having little or no in-domain manually transcribed training data;

operating the semantic classification system to assign input speech utterances to the defined semantic classifications;

collecting adaptation training data based on input speech utterances with manually assigned semantic labels; and

when the adaptation training data satisfies a pre-determined adaptation criteria, automatically retraining the semantic classification system based on the adaptation training data.

2. A method according to claim 1, wherein training the semantic classification system includes training a language model in an automatic speech recognition engine that converts the input speech utterances into representative text.

3. A method according to claim 1, wherein training the semantic classification system includes training a statistical semantic classifier that classifies text representing the input speech utterances according to one of the defined semantic classifications.

4. A method according to claim 1, wherein training the semantic classification system includes training a confidence engine that determines confidence scores based on acoustic features and natural language understanding (NLU) features to represent the likelihood of the input speech utterances being correctly classified with a defined semantic classification.

5. A method according to claim 1, wherein the training data includes in-domain automatically transcribed training data.

6. A method according to claim 1, wherein the training data includes manually transcribed training data from one or more other semantic classification systems.

7. A method according to claim 1, wherein the training data includes a small amount of in-domain manually transcribed training data.

8. A method according to claim 1, wherein the semantic classification system is a call routing system.

9. A method of automated training for a language model in a semantic classification system, the method comprising:

performing first-pass recognition of available training data using an initial language model to develop first pass recognition results corresponding to the available training data;

dividing the first pass recognition results into a plurality of first pass subsets;

for each first pass subset, developing a subset language model from the first pass subset recognition results;

performing second-pass recognition of the first pass subsets using cross-adapted subset language models to develop second pass recognition results corresponding to the available training data; and

computing a final adapted language model from the second pass recognition results.

10. A method according to claim 9, wherein the first pass recognition results are divided into two first pass subsets.

11. A method according to claim 9, wherein the first pass subsets are approximately equal in total duration.

12. A method according to claim 9, further comprising:

computing a statistical semantic classifier from the second pass recognition results for assigning input speech utterances to predefined call routes.

13. A method of automated training for a confidence engine in a semantic classification system, the method comprising:

operating an under-trained semantic classification system on a set of input speech utterances with manually assigned semantic labels to extract a set of corresponding feature data; and

training the confidence engine on the extracted feature data.

14. A method of automated training for a confidence engine in a semantic classification system, the method comprising:

operating a semantic classification system on a set of input speech utterances with manually assigned semantic labels to extract a set of corresponding feature data; and

training on the extracted feature data the confidence engine for a different semantic classification application.

15. A method of automated training for a router confidence engine in a call routing system, the method comprising:

defining a set of call routes for processing input speech utterances in a call routing system, each call route representing a specific call processing path;

performing initial router confidence training based on an initial set of training data;

operating the call routing system to assign input speech utterances to the defined call routes;

collecting adaptation training data based on processed input speech utterances; and

re-performing router confidence training based on the adaptation training data.

16. A method of training for an automated call routing system for improving system accuracy at early deployment stages, the method comprising:

defining a set of call routes for processing input speech utterances in a call routing system, each call route representing a specific call processing path;

performing a first training of the call routing system from training data having little or no in-domain manually transcribed training data and a large amount of un-transcribed in-domain data;

operating the call routing system to assign input speech utterances to the defined call routes;

collecting adaptation training data based on manually processed input speech utterances; and

automatically retraining the call routing system based on the adaptation training data to improve system accuracy.

17. A method of training for an automated call routing system for improving system accuracy at early deployment stages, the method comprising:

defining a set of call routes for processing input speech utterances in a call routing system, each call route representing a specific call processing path;

training the call routing system based on a generic confidence engine;

operating the call routing system to assign input speech utterances to the defined call routes;

collecting adaptation training data based on manually processed input speech utterances; and

automatically retraining the call routing system based on the adaptation training data to improve system accuracy.

18. A method of training for an automated call routing system for improving system accuracy at early deployment stages, the method comprising:

defining a set of call routes for processing input speech utterances in a call routing system, each call route representing a specific call processing path;

training the call routing system based on a generic confidence engine using little or no in-domain manually transcribed training data and a large amount of un-transcribed in-domain data;

operating the call routing system to assign input speech utterances to the defined call routes;

collecting adaptation training data based on manually processed input speech utterances; and

automatically retraining the call routing system based on the adaptation training data to improve system accuracy.