Static analysis to identify defects in grammars

• US 7,711,551 B2
• Filed: 06/13/2005
• Issued: 05/04/2010
• Est. Priority Date: 06/13/2005
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Campbell, Craig, Lopez-Barquilla, Ricardo
• Primary Examiner(s): Hudspeth; David R
• Assistant Examiner(s): Rider; Justin W
• Application Number:
  US11/151,576
• Publication Number:
  US 20060282267A1
• Time in Patent Office: 1,786 Days
• Field of Search: 704/9, 704/270.1
• US Class Current: 704/9
• CPC Class Codes: G10L 15/18 using natural language modelling
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models

Abstract

The present invention provides static analysis of speech grammars prior to the speech grammars being deployed in a speech system.

Claims

1. A method of analyzing a context free grammar for a speech system during authoring of the context free grammar, comprising:

authoring at least a portion of the context free grammar;

loading the authored context free grammar into a grammar analyzer having a plurality of static analysis components each capable of identifying a different class of grammar defects, wherein at least one of the classes is related to identifying defects related to speech recognition and at least one of the classes is related to identifying defects related to something other than speech recognition;

intermittently, while authoring the context free grammar, selecting one of the plurality of accessible static analysis components;

running the selected static analysis component on the context free grammar to identify specific defects in the context free grammar of the class associated with the selected static analysis component wherein the selected static analysis component identifies over-generation defects in the context free grammar by selecting each text fragment allowed by the context free grammar, calculating a language model score for the selected text fragment with a language model to determine whether the selected text fragment is likely to be used by a user based on the language model score; and

if not, generating an output indicative of a possibility that an unusual text fragment, allowed by the context free grammar, has been identified; and

repeating the steps of selecting and running for each desired static analysis component.

2. The method of claim 1 wherein determining whether the text fragment is likely to be used comprises:

parsing an input with the context free grammar to build the text fragment; and

calculating the score for the built text fragment using the language model.

3. The method of claim 2 wherein determining whether the text fragment is likely to be used further comprises:

comparing the score to a threshold value; and

wherein the step of generating an output indicative of a possibility that an unusual text fragment, allowed by the context free grammar, has been identified if the score does not meet the threshold value.

4. The method of claim 1 wherein the selected static analysis component identifies acoustically confusable tokens in the context free grammar.

5. The method of claim 4 wherein identifying acoustically confusable tokens comprises:

identifying a token in the context free grammar;

generating synthetic audio information indicative of a pronunciation of the identified token; and

determining whether the synthetic audio information is acoustically confusable with acoustic information corresponding to other tokens in the context free grammar.

6. The method of claim 5 wherein determining whether the synthetic audio information is acoustically confusable with acoustic information corresponding to other tokens in the context free grammar comprises:

performing speech recognition on the synthetic audio information to obtain a speech recognition result that includes a best speech recognition result and one or more alternate speech recognition results; and

determining whether the best speech recognition result is too close to the one or more alternate speech recognition results.

7. The method of claim 5 wherein determining whether the synthetic audio information is acoustically confusable with acoustic information corresponding to other tokens in the context free grammar comprises:

perturbing the synthetic audio information;

performing speech recognition on the perturbed audio information to obtain a speech recognition result; and

determining whether the synthetic audio information is acoustically confusable with acoustic information corresponding to other tokens in the context free grammar based on the speech recognition result.

8. The method of claim 1 wherein the selected static analysis component identifies spelling and pronunciation defects in the context free grammar.

9. The method of claim 8 wherein identifying spelling and pronunciation defects comprises:

identifying a word in the context free grammar; and

accessing a lexicon with words and corresponding pronunciations; and

determining whether the identified word is misspelled or has an incorrect pronunciation based on the lexicon.

10. The method of claim 9 wherein determining whether the identified word is misspelled or has an incorrect pronunciation comprises:

if the word in the context free grammar has a pronunciation specified in the context free grammar, then comparing the specified pronunciation to a pronunciation in the lexicon corresponding to the word; and

if the distance between the specified pronunciation and the pronunciation in the lexicon meets a threshold distance, then providing an indication that the spelling or pronunciation may be in error.

11. The method of claim 9 wherein determining whether the identified word is misspelled or has an incorrect pronunciation comprises:

if the word in the context free grammar does not have a pronunciation specified in the context free grammar, then determining whether the word is in the lexicon; and

if the word is not in the lexicon, then providing an indication that the spelling of the word may be in error.

12. The method of claim 11 wherein providing an indication that the spelling of the word may be in error comprises:

determining whether the word is in a frequently misspelled words data store; and

if the word is in the frequently misspelled words data store, then providing a suggested correct spelling from the frequently misspelled words data store.

13. A static analyzer that includes a processor capable of executing instructions for analyzing a speech context free grammar, comprising:

an authoring component used to author the context free grammar, including an analysis component performing static analysis on the context free grammar by executing instructions in the processor for selecting one or more defect identifying components, with at least one of the defect identifying components being related to speech recognition defects and at least one of the defect identifying components being related to something other than speech recognition components and running the selected defect identifying components on the context free grammar to identify specific defects in the context free grammar, wherein the defect identifying components include a syntax error identifying component that identifies syntax errors in the context free grammar, the syntax errors including an external grammar defect identified when the syntax error identifying component determines that references, in the context free grammar, to external rules cannot be resolved or that the external rules cannot be loaded into the analysis component; and

wherein the defect identifying components comprise a specific engine defect identifier identifying whether the context free grammar violates engine specific limits for a plurality of different speech recognition engines, each of the engine specific limits corresponding to one of the plurality of different speech recognition engines.

14. The static analyzer of claim 13 wherein the defect identifying components comprise an over-generation defect identifier identifying over-generation in the context free grammar by determining whether text fragments allowed by the context free grammar are infrequently used text fragments.

15. The static analyzer of claim 13 wherein the defect identifying components comprise an acoustic defect identifier identifying whether tokens in the context free grammar are undesirably acoustically similar to one another based on synthetic audio information generated for at least one of the tokens.

16. The static analyzer of claim 13 wherein the defect identifying components comprise a spelling and pronunciation defect identifier identifying spelling and pronunciation defects in the context free grammar by accessing information in a lexicon to determine whether the information in the context free grammar is similar to the information in the lexicon.