Adapting a language model to accommodate inputs not found in a directory assistance listing

• US 8,285,542 B2
• Filed: 02/15/2011
• Issued: 10/09/2012
• Est. Priority Date: 12/19/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Yu, Dong, Acero, Alejandro, Ju, Yun-Cheng
• Primary Examiner(s): Han, Qi
• Application Number:
  US13/027,921
• Publication Number:
  US 20110137639A1
• Time in Patent Office: 602 Days
• Field of Search: 704/1, 704/9, 704/10, 704/231, 704/251, 704/257, 704/235, 704/255
• US Class Current: 704/10
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams

Abstract

A statistical language model is trained for use in a directory assistance system using the data in a directory assistance listing corpus. Calculations are made to determine how important words in the corpus are in distinguishing a listing from other listings, and how likely words are to be omitted or added by a user. The language model is trained using these calculations.

Claims

1. A method of generating a language model from a corpus of directory assistance listings, the language model used to estimate a desired directory assistance listing based on a listing request input, the method comprising:

for each word in each listing in the corpus, generating, using a processing unit, language model counts attributable to the word by;

determining, based on one or more characteristics of the word in the listing, whether the word is likely to be omitted from the listing request input when the desired directory assistance listing is a listing that contains the word;

generating the language model counts attributable to the word based on whether the word is likely to be omitted from the listing request input when the desired directory assistance listing is a listing that contains the word; and

generating the language model counts attributable to the word based on a characteristic indicative of how well the word operates to distinguish the listing that contains the word from other listings in the corpus; and

storing the language model with the language model counts.

2. The method of claim 1 wherein the corpus is divided into categories of directory assistance listings, and further comprising:

generating the language model counts attributable to the word based on an indicator of whether the word is descriptive of the category to which the listing that contains the word belongs.

3. The method of claim 2 wherein generating the language model counts attributable to the word comprises:

generating the language model counts attributable to the word based on whether the word is likely to be omitted, generating the language model counts attributable to the word based on a characteristic indicative of how well the word operates to distinguish the listing that contains the word from other listings in the corpus, and generating the language model counts attributable to the word based on an indicator of whether the word is descriptive of the category to which it belongs, separately for words in each category in the corpus.

4. The method of claim 3 wherein generating the language model counts attributable to the word based on whether the word is likely to be omitted comprises calculating a probability that the word will be omitted from the listing request input.

5. The method of claim 2, wherein each category has a textual label, and the method comprises:

generating the language model counts attributable to the word based on whether the word is in the textual label for the category that contains the listing to which the word belongs.

6. The method of claim 1 wherein the one or more characteristics of the word in the listing comprises a word position of the word in the listing that contains the word.

7. The method of claim 1, wherein the corpus of directory assistance listings is divided into categories, and the one or more characteristics comprise an irriportance measure based on a frequency of occurrence of the word in the category to which the listing, that contains the word, belongs.

8. The method of claim 7, wherein the importance measure is calculated based on a term frequency-inverse document frequency value for the word in the category.

9. The method of claim 7, wherein the importance measure is calculated based on a maximum mutual information value for the word in the category.

10. A system for generating a language model used to estimate a desired directory assistance listing based on a listing request input, the system comprising:

a processor;

a corpus of directory assistance listings; and

a language model generation component configured to;

access the corpus of directory assistance listings;

for each word in each listing in the corpus,

determine, based on one or more characteristics of the word in the listing, whether the word is likely to be omitted from the listing request input when the desired directory assistance listing is a listing that contains the word;

generate language model counts attributable to the word using the processor, wherein the language model counts are generated based on whether the word is likely to be omitted from the listing request input when the desired directory assistance listing is a listing that contains the word, and based on a characteristic indicative of how well the word operates to distinguish the listing that contains the word from other listings in the corpus; and

store the language model with the language model counts.

11. The system of claim 10, wherein the corpus is divided into categories of directory assistance listings, and wherein the language model generation component is configured to generate the language model counts attributable to the word based on an indicator of whether the word is descriptive of the category to which the listing that contains the word belongs.

12. The system of claim 11, wherein the language model generation component is configured to generate, separately for words in each category in the corpus:

the language model counts attributable to the word based on whether the word is likely to be omitted;

the language model counts attributable to the word based on a characteristic indicative of how well the word operates to distinguish the listing that contains the word from other listings in the corpus; and

the language model counts attributable to the word based on an indicator of whether the word is descriptive of the category to which it belongs.

13. The system of claim 12, wherein the language model generation component is configured to generate the language model counts attributable to the word based on whether the word is likely to be omitted comprises calculating a probability that the word will be omitted from the listing request input.

14. The system of claim 11, wherein each category has a textual label, and the language model generation component is configured to:

generate the language model counts attributable to the word based on whether the word is in the textual label for the category that contains the listing to which the word belongs.

15. The system of claim 10, wherein the one or more characteristics of the word in the listing comprises a word position of the word in the listing that contains the word.

16. The system of claim 10, wherein the corpus of directory assistance listings is divided into categories, and the one or more characteristics comprise an importance measure based on a frequency of occurrence of the word in the category to which the listing, that contains the word, belongs.

17. The system of claim 16, wherein the importance measure is calculated based on a term frequency-inverse document frequency value for the word in the category.

18. The system of claim 16, wherein the importance measure is calculated based on a maximum mutual information value for the word in the category.