Method and system for using rule-based knowledge to build a class-based domain specific statistical language model

• US 7,275,033 B1
• Filed: 09/30/2000
• Issued: 09/25/2007
• Est. Priority Date: 09/30/2000
• Status: Expired due to Fees

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Yan, Yonghong, Lin, Zhiwei, Zhao, Yibao
• Primary Examiner(s): Hudspeth; David
• Assistant Examiner(s): Sked; Matthew J.
• Application Number:
  US10/130,860
• Time in Patent Office: 2,551 Days
• Field of Search: None
• US Class Current: 704/255
• 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

A method and system for providing a class-based statistical language model representation from rule-based knowledge is disclosed. The class-based language model is generated from a statistical representation of a class-based rule net. A class-based rule net is generated using the domain-related rules with words replaced with their corresponding class-tags that are manually defined. The class-based statistical representation from the class-based rule net is combined with a class-based statistical representation from a statistical language model to generate a language model. The language model is enhanced by smoothing/adapting with general-purpose and/or domain-related corpus for use as the final language model. A two-pass search algorithm is applied for speech decoding.

Claims

1. A speech processing method comprising:

manually defining a set of classes for words/phrases, each class being associated with a class-tag;

replacing words of domain-related rules with corresponding class-tags to generate class-based domain-related rules;

generating a class-based rule net using the class-based domain-related rules;

obtaining a class-based statistical representation from the class-based rule net; and

combining the class-based statistical representation with a class-based statistical representation from a statistical language model to generate a final language model.

2. The speech processing method of claim 1, wherein replacing words comprising:

selecting words with same function and semantic meaning;

associating each of the words with one of the set of classes; and

replacing each of the words with a class-tag of class that is associated with the word in the domain-related rules.

3. The speech processing method of claim 1, further comprising applying a two-pass search method for decoding speech the final language model.

4. The speech processing method of claim 3, wherein the two-pass search method comprises:

obtaining a sentence clause from a rule net using a first pass search process; and

obtaining a final result from obtained sentence clause using a second pass search process.

5. The speech processing method of claim 4, wherein the first pass search process further includes:

determining if a word belongs to a predefined class;

associating the determined word with the predefined class; and

creating a new class for the word that does not belong to a predefined class.

6. The speech processing method of claim 1, wherein combining the class-based statistical representations comprises:

using a domain-related and/or general-purpose corpus;

preprocessing the domain-related and/or general-purpose corpus to replace words in the domain-related and/or general-purpose corpus with corresponding class-tags to generate a class-based corpus;

obtaining a class-based statistical representation from the class-based corpus; and

merging the class-based statistical representation from the class-based corpus with the class-based statistical representation from the class-based rule net to generate the final language model.

7. The speech processing method of claim 1, further comprising applying smoothing/adapting to the final language model before using as a final language model.

8. A system comprising:

a memory to store a statistical language model representation of a class-tag based rule net; and

a processor to enable a user to manually define a set of classes for words/phrases, each class being associated with a class-tag, to replace words of domain-related rules with corresponding class-tags to generate class-based domain-related rules, to generate a class-based rule net using the class-based domain-related rules, to obtain a class-based statistical representation from the class-based rule net, and to combine the class-based statistical representation with a class-based statistical representation from a statistical language model to generate a final language model.

9. The system of claim 8, wherein the processor replaces the words with class-tags by selecting words with same function and semantic meaning, associating each of the words with one of the set of classes, and replacing each of the words with a class-tag of class that is associated with the word in the domain-related rules.

10. The system of claim 8, wherein the processor further applies a two-pass search method for decoding speech the final language model.

11. The system of claim 10, wherein the two-pass search method comprises obtaining a sentence clause from a rule net using a first pass search method, and obtaining a final result from obtained sentence clause using a second pass search method.

12. The system of claim 11, wherein the processor is to apply the first pass search process by determining if a word belongs to a predefined class, to associate the determined word with the predefined class, and to create a new class for the word that does not belong to a predefined class.

13. The system of claim 8, wherein the processor combines the class-based statistical representations by:

using a domain-related and/or general-purpose corpus;

preprocessing the domain-related and/or general-purpose corpus to replace words in the domain-related and/or general-purpose corpus with corresponding class-tags to generate a class-based corpus; and

obtaining a class-based statistical representation from the class-based corpus; and

merging the class-based statistical representation from the class-based corpus with the class-based statistical representation from the class-based rule net to generate the final language model.

14. The system of claim 8, wherein the processor is to apply smoothing/adapting to the final language model before using as a final language model.

15. A machine readable storage medium having stored therein a plurality of machine executable instructions, wherein when executed, the instructions perform operations comprising:

manually defining a set of classes for words/phrases, each class being associated with a class-tag;

replacing words of domain-related rules with corresponding class-tags to generate class-based domain-related rules;

generating a class-based rule net using the class-based domain-related rules;

obtaining a class-based statistical representation from the class-based rule net; and

combining the class-based statistical representation with a class-based statistical representation from a statistical language model to generate a final language model.

16. The storage medium of claim 15, wherein replacing words further includes:

selecting words with same function and semantic meaning;

associating each of the words with one of the set of classes; and

replacing each of the words with a class-tag of class that is associated with the word in the domain-related rules.

17. The storage medium of claim 15, wherein the operations further comprise applying a two-pass search method for decoding speech with the final language model.

18. The storage medium of claim 17, wherein the two-pass search method comprises:

obtaining a sentence clause from a rule net using a first pass search process; and

obtaining a final result from obtained sentence clause using a second pass search method.

19. The storage medium of claim 18, wherein the first pass search process further includes:

determining if a word belongs to a predefined class;

associating the determined word with the predefined class; and

creating a new class for a word that does not belong to a predefined class.

20. The storage medium of claim 15, wherein combining the class-based statistical representations comprises:

using a domain-related and/or general-purpose corpus;

preprocessing the domain-related and/or general-purpose corpus to replace words in the domain-related and/or general-purpose corpus with corresponding class-tags to generate a class-based corpus;

obtaining a class-based statistical representation from the class-based corpus; and

merging the class-based statistical representation from the class-based corpus with the class-based statistical representation from the class-based rule net to generate the final language model.

21. The storage medium of claim 15, wherein the operations further comprises applying smoothing/adapting to the final language model before using as a final language model.