LANGUAGE MODEL CREATION DEVICE, LANGUAGE MODEL CREATION METHOD, AND COMPUTER-READABLE STORAGE MEDIUM

• US 20120035915A1
• Filed: 03/16/2010
• Published: 02/09/2012
• Est. Priority Date: 04/30/2009
• Status: Active Grant

• Current Assignee: NEC Corporation
• Original Assignee: NEC Corporation
• Inventors: Kitade, Tasuku, Koshinaka, Takafumi, Onishi, Yoshifumi
• Granted Patent: US 8,788,266 B2
• Time in Patent Office: Days
• US Class Current: 704/9
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams

Abstract

The present invention uses a language model creation device 200 that creates a new language model using a standard language model created from standard language text. The language model creation device 200 includes a transformation rule storage section 201 that stores transformation rules used for transforming dialect-containing word strings into standard language word strings, and a dialect language model creation section 203 that creates dialect-containing n-grams by applying the transformation rules to word n-grams in the standard language model and, furthermore, creates the new language model (dialect language model) by adding the created dialect-containing n-grams to the word n-grams.

Claims

1. A language model creation device that creates a new language model using a standard language model created from standard language text, comprising:

a transformation rule storage section that stores transformation rules used for transforming dialect-containing word strings into standard language word strings, and

a dialect language model creation section that creates dialect-containing n-grams by applying the transformation rules to word n-grams in the standard language model and, furthermore, creates the new language model by adding the created dialect-containing n-grams to the word n-grams.

2. The language model creation device according to claim 1,

wherein the transformation rule storage section stores groups each made up of a dialect-containing word string and a standard language-containing word string corresponding to the dialect as the transformation rules, and

the dialect language model creation section retrieves the probability of occurrence of the standard language-containing word string from the standard language model and computes the probability of occurrence of the dialect-containing n-grams based on the retrieved probability of occurrence and a pre-configured distribution probability.

3. The language model creation device according to claim 2,

wherein the dialect language model creation section configures a value of the distribution probability using dialect data including dialect-containing speech data and dialect-containing text data.

4. The language model creation device according to claim 2, further comprising a transformation rule processing section that extracts dialect-containing word strings from the dialect data and modifies the transformation rules on the basis of the extracted dialect-containing word strings, and

wherein the transformation rule storage section updates the already stored transformation rules using the transformation rules modified by the transformation rule processing section.

5. The language model creation device according to claim 2, further comprising a transformation rule processing section that extracts dialect-containing word strings from the dialect data and derives transformation patterns that can be utilized as the transformation rules using the extracted dialect-containing word strings and standard language word strings corresponding to the extracted dialect-containing word strings.

6. The language model creation device according to claim 4,

wherein, in case no transformation rules are stored in the transformation rule storage section, the transformation rule processing section extracts, from the dialect data, word strings that are not included in the standard language model among the word strings contained therein and creates the transformation rules using the extracted word strings.

7. The language model creation device according to claim 3,

wherein after configuring the distribution probability value, creating the dialect-containing n-grams, and creating the new language model, the dialect language model creation section acquires speech recognition results obtained by an external speech recognition device utilizing the new language model along with the dialect data and updates the distribution probability value based on the acquired speech recognition results and ground truth data used for the dialect data.

8. The language model creation device according to claim 1, further comprising:

a transformed data creation section that extracts the dialect-containing word strings from dialect data including the dialect-containing speech data and the dialect-containing text data and uses the transformation rules to transform the extracted dialect-containing word strings into word strings containing only standard language;

a transformed data storage section that stores, as transformed data, the word strings containing only standard language obtained via transformation by the transformed data creation section;

a standard language data storage section that stores standard language text data; and

a standard language model creation section that creates the standard language model using the transformed data stored by the transformed data storage section and the standard language text data stored by the standard language data storage section.

9. The language model creation device according to claim 8,

wherein the data creation section extracts dialect-containing word strings from the dialect data and modifies the transformation rules on the basis of the extracted dialect-containing word strings, and

the transformation rule storage section updates the already stored transformation rules using the transformation rules modified by the transformation rule processing section.

10. The language model creation device according to claim 8,

wherein the transformed data creation section extracts dialect-containing word strings from the dialect data and derives transformation patterns that can be utilized as the transformation rules using the extracted dialect-containing word strings and standard language word strings corresponding to the extracted dialect-containing word strings.

11. The language model creation device according to claim 9,

wherein;

in case no transformation rules are stored in the transformation rule storage section, the transformed data generation section extracts, from the dialect data, word strings that are not included in the standard language model among the word strings contained therein and creates the transformation rules using the extracted word strings.

12. A language model creation method for creating a new language model using a standard language model created from standard language text, the method comprising:

(a) configuring transformation rules used for transforming dialect-containing word strings into standard language word strings, and

(b) creating dialect-containing n-grams by applying the transformation rules to word n-grams in the standard language model and, furthermore, creating a new language model by adding the created dialect-containing n-grams to the word n-grams.

13. The language model creation method according to claim 12,

wherein the configuring includes configuring groups each made up of a dialect-containing word string and a standard language-containing word string corresponding to the dialect as the transformation rules, and

the creating, after creating the dialect-containing n-grams, includes retrieving the probability of occurrence of the standard language-containing word string from the standard language model and computing the probability of occurrence of the added dialect-containing word string of the same group based on the retrieved probability of occurrence and a pre-configured distribution probability.

14. The language model creation method according to claim 13,

wherein the creating includes configuring the distribution probability value using dialect data including the dialect-containing speech data and the dialect-containing text data.

15. The language model creation method according to claim 13, further comprising:

(c) extracting dialect-containing word strings from the dialect data and modifying the transformation rules on the basis of the extracted dialect-containing word strings, and

(d) updating the transformation rules already configured in the configuring using the transformation rules modified in the extracting.

16. The language model creation method according to claim 13, further comprising (e) extracting dialect-containing word strings from the dialect data, extracting standard language word strings corresponding to the extracted dialect-containing word strings from the standard language model, and deriving transformation patterns that can be utilized as the transformation rules using the extracted dialect-containing word strings and the extracted standard language word strings.

17. The language model creation method according to claim 14, further comprising (f) after configuring the distribution probability value, creating the dialect-containing n-grams, and creating the new language model in the creating, acquiring speech recognition results obtained by an external speech recognition device utilizing the new language model and the dialect data and updating the distribution probability value used in the creating, based on the acquired speech recognition results and the ground truth data of the dialect data.

18. The language model creation method according to claim 12, further comprising:

(g) extracting the dialect-containing word strings from the dialect data including the dialect-containing speech data and the dialect-containing text data and using the transformation rules to transform the extracted dialect-containing word strings to word strings containing only standard language, and

(h) creating the standard language model using standard language text data and the word strings containing only standard language transformed in the extracting.

19. A computer-readable storage medium having recorded thereon a software program used for creating a new language model utilizing a standard language model created from standard language text on a computer, the software program comprising instructions directing the computer to perform:

(a) configuring transformation rules used for transforming dialect-containing word strings into standard language word strings, and

(b) creating dialect-containing n-grams by applying the transformation rules to word n-grams in the standard language model and, furthermore, creating the new language model by adding the created dialect-containing n-grams to the word n-grams.

20. The computer-readable storage medium according to claim 19,

wherein the configuring includes configuring groups each made up of a dialect-containing word string and a standard language-containing word string corresponding to the dialect as the transformation rules, and

the creating, after creating the dialect-containing n-grams, includes retrieving the probability of occurrence of the standard language-containing word string from the standard language model and computing the probability of occurrence of the added dialect-containing word string of the same group based on the retrieved probability of occurrence and a pre-configured distribution probability.

21. The computer-readable storage medium according to claim 20,

wherein the creating includes configuring the distribution probability value using dialect data including the dialect-containing speech data and the dialect-containing text data.

22. The computer-readable storage medium according to claim 20, the software program further comprising instructions directing the computer to execute:

(c) extracting dialect-containing word strings from the dialect data and modifying the transformation rules on the basis of the extracted dialect-containing word strings, and

(d) updating the transformation rules already configured in the configuring using the transformation rules modified in the step (c).

23. The computer-readable storage medium according to claim 20, the software program further comprising instructions directing the computer to execute (e) extracting dialect-containing word strings from the dialect data, extracting standard language word strings corresponding to the extracted dialect-containing word strings from the standard language model, and deriving transformation patterns that can be utilized as the transformation rules using the extracted dialect-containing word strings and the extracted standard language word strings.

24. The computer-readable storage medium according to claim 19, the software program further comprising instructions directing the computer to execute (f) after configuring the distribution probability value, creating the dialect-containing n-grams, and creating the new language model in the creating, acquiring speech recognition results obtained by an external speech recognition device utilizing the new language model and the dialect data and updating the distribution probability value used in the creating based on the acquired speech recognition results and the ground truth data of the dialect data.

25. The computer-readable storage medium according to claim 19, the software program further comprising instructions directing the computer to execute:

(g) extracting the dialect-containing word strings from the dialect data including the dialect-containing speech data and the dialect-containing text data and using the transformation rules to transform the extracted dialect-containing word strings to word strings containing only standard language, and

(h) creating the standard language model using standard language text data and the word strings containing only standard language transformed in the extracting.