Disambiguation language model

• US 7,251,600 B2
• Filed: 03/29/2005
• Issued: 07/31/2007
• Est. Priority Date: 01/31/2001
• Status: Expired due to Fees

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Ju, Yun-cheng, Alleva, Fileno
• Primary Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US11/092,252
• Publication Number:
  US 20050171761A1
• Time in Patent Office: 854 Days
• Field of Search: 704/257, 704/8, 704/9, 704/10, 704/270
• US Class Current: 704/257
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/18 using natural language modelling

Abstract

A language model for a language processing system such as a speech recognition system is constructed from training corpus formed from associated characters, word phrases and context cues. A method and apparatus for generating the training corpus used to train the language model and a system or module using such a language model is disclosed.

Claims

1. A computer readable medium including instructions readable by a computer which, when implemented, cause the computer to construct a language model for a speech recognition system to disambiguate characters of an Asian language, the method comprising constructing a training corpus comprising the steps of:

obtaining a dictionary of word phrases;

for each word phrase of the dictionary of word phrases comprising Kanji-based characters, associating a character of the word phrase and the word phrase with a context cue indicative of disambiguating the character to automatically generate context cue phrases of the training corpus; and

using the training corpus to build the language model.

2. The computer readable medium of claim 1, wherein the language model comprises an N-gram language model having probability information for the generated phrases.

3. The computer readable medium of claim 1, wherein associating includes associating a first or last character of each word phrase with the word phrase.

4. The method of claim 2, and further comprising adjusting a probability score for each of the associated characters and word phrases in the language model.

5. The method of claim 1, wherein the context cue comprises in Chinese.

6. The method of claim 1, wherein the context cue comprises in Japanese.

7. The method of claim 1, wherein each character is a single Kanji-based character.

8. A computer readable medium including instructions which, when implemented, cause a computer to recognize Kanji-based characters when spoken, the instructions comprising:

a first module adapted to construct a training corpus by performing the steps;

obtaining a dictionary of word phrases;

for each word phrase of the dictionary of word phrases comprising Kanji-based characters, associating a character of the word phrase and the word phrase with a context cue indicative of disambiguating the character to automatically generate context cue phrases of the training corpus; and

a second module adapted to construct a language model using the training corpus.

9. The computer readable medium of claim 8, wherein associating includes associating a first or last character of each word phrase with the word phrase.

10. The computer readable medium of claim 9, wherein the context cue comprises in Chinese.

11. The method of claim 9, wherein the context cue comprises in Japanese.

12. The computer readable medium of claim 8, and further comprising a third module adapted to recognize Kanji-based characters when spoken by performing the step of receiving input speech having a context cue phrase, the context cue phrase comprising a Kanji-based character, a word phrase having the Kanji-based character, and a context cue, wherein the context cue is indicative of disambiguating the Kanji-based character.

13. The computer readable medium of claim 12, wherein the third module is further adapted to perform the steps of:

detecting the context cue phrase in the received input speech without prompting;

executing instructions for accessing the language model,

wherein the language model comprises a statistical language model having probability information for the context cue phrases; and

outputting the character as text without the word phrase and the context cue for the detected context cue phrase.

14. The computer readable medium of claim 13, wherein the statistical language model is an N-gram language model.

15. A method of recognizing Kanji-based characters when spoken, the method comprising the steps of:

receiving input speech having a context cue phrase, the context cue phrase comprising a Kanji-based character, a word phrase having the Kanji-based character, and a context cue,

wherein the context cue is indicative of disambiguating the Kanji-based character;

detecting the context cue phrase in the received input speech without prompting;

executing instructions for accessing a language model, wherein the language model comprises an N-gram language model having probability information for the context cue phrases; and

outputting the character as text without the word phrase and the context cue for the detected context cue phrase.

16. The method of claim 15, wherein outputting the character includes outputting the character string based on recognizing the character using the language model.

17. The method of claim 15, wherein the language model comprises a context-free grammar.

18. The method of claim 15, wherein outputting the character includes outputting the character based on a comparison of a recognized character with a recognized word phrase.

19. The method of claim 18, wherein when the recognized character is not present in the recognized word phrase, the character that is output is a character of the recognized word phrase.

20. The method of claim 15, wherein each of the word phrases is a single word, and wherein each of the characters is a single Kanji-based character.