Dynamically adding personalization features to language models for voice search

• US 8,938,391 B2
• Filed: 06/12/2011
• Issued: 01/20/2015
• Est. Priority Date: 06/12/2011
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Zweig, Geoffrey, Chang, Shuangyu
• Primary Examiner(s): Vo, Huyen X.
• Application Number:
  US13/158,453
• Publication Number:
  US 20120316877A1
• Time in Patent Office: 1,318 Days
• Field of Search: 704 1- 10, 704/251, 704/255, 704/257, 704/270, 704/270.1, 704/231, 704/235
• US Class Current: 704/251
• CPC Class Codes: G10L 15/07 to the speaker
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams

Abstract

A dynamic exponential, feature-based, language model is continually adjusted per utterance by a user, based on the user'"'"'s usage history. This adjustment of the model is done incrementally per user, over a large number of users, each with a unique history. The user history can include previously recognized utterances, text queries, and other user inputs. The history data for a user is processed to derive features. These features are then added into the language model dynamically for that user.

Claims

1. A computer system for voice search comprising:

first storage in which data defining an exponential language model are stored for use for a plurality of users;

second storage in which data defining user histories are stored, with a user history for each user of the plurality of users, the user history for each user storing data describing utterances processed for the user; and

one or more processing devices connected to access the first storage and the second storage, and in communication with a plurality of user devices to receive and process utterances received from the plurality of users through the plurality of user devices, each of the received utterances being provided by a user of the plurality of users, the one or more processing devices configured to;

for each user of the plurality of users, define personalization features for the user from the user history of the user in the second storage, the personalization features corresponding to features in the exponential language model; and

for each of the received utterances;

process the utterance using both the features in the exponential language model from the first storage and the personalization features defined from the user history of the user providing the utterance to identify words in the utterance;

provide search results using the words identified in the utterance to the user device from which the utterance was received; and

update, in the second storage, the user history of the user providing the utterance based on the utterance received from the user.

2. The computer system of claim 1, wherein the one or more processing devices is further configured to continually update the user histories of each user in the second storage according to utterances from each user processed by the processing device and to update the personalization features for each user based on the updated user histories in the second storage.

3. The computer system of claim 1, wherein personalization features of a same size have a same weight.

4. The computer system of claim 1, wherein the personalization features include word based features including words from the user history.

5. The computer system of claim 1, wherein the personalization features include class based features including classes of words from the user history.

6. The computer system of claim 1, wherein the personalization features include features with both words and class information.

7. The computer system of claim 1, wherein the personalization features include information derived from a locality of the user and words associated with the locality.

8. A computer-implemented process, comprising:

maintaining in a first storage accessible by one or more server computers data defining an exponential language model for use for a plurality of users;

maintaining in a second storage accessible by the one or more server computers data defining user histories, with a user history for each user of the plurality of users, the user history for each user storing data describing utterances processed for the user;

establishing communication between the one or more server computers and a plurality of user devices associated with the plurality of users, to enable the server computers to receive and process utterances received from the plurality of users through the plurality of user devices, each of the received utterances being provided by a user of the plurality of users;

the one or more server computers defining, for each of the plurality of users, personalization features for the user from the user history of the user in the second storage, the personalization features corresponding to features in the exponential language model; and

in response to each of the utterances received by the one or more server computers from any user of the plurality of users, the one or more server computers;

processing the utterance using both the features in the exponential language model from the first storage and the personalization features defined from the user history of the user providing the utterance to identify words in the utterance;

providing search results using the words identified in the utterance to the user device from which the utterance was received; and

updating, in the second storage, the user history of the user providing the utterance in based on the utterance received from the user.

9. The computer implemented process of claim 8, further comprising continually updating the user histories of each user in the second storage according to utterances from each user processed by the one or more server computers and to update the personalization features for each user based on the updated user histories in the second storage.

10. The computer implemented process of claim 8, wherein personalization features of a same size have a same weight.

11. The computer implemented process of claim 8, wherein the personalization features include word based features including words from the user history.

12. The computer implemented process of claim 8, wherein the personalization features include class based features including classes of words from the user history.

13. The computer implemented process of claim 8, wherein the personalization features include features with both words and class information.

14. The computer implemented process of claim 8, wherein the personalization features include information derived from a locality of the user and words associated with the locality.

15. An article of manufacture, comprising:

a computer-readable storage medium;

computer program instructions encoded on the computer-readable storage medium that, when processed by a computer, instruct the computer to perform a process comprising;

maintaining, in a first storage accessible by the computer, data defining an exponential language model for use for a plurality of users;

maintaining, in a second storage accessible by the computer, data defining user histories, with a user history for each user of the plurality of users, the user history for each user storing data describing utterances processed for the user;

enabling the computer to receive utterances from a plurality of user devices associated with the plurality of users;

the computer defining, for each user of the plurality of users, personalization features for the user from the user history of the user in the second storage, the personalization features corresponding to features in the exponential language model; and

for each of the received utterances;

processing the utterance using both the features in the exponential language model from the first storage and the personalization features defined from the user history of the user providing the utterance to identify words in the utterance;

providing search results using the words identified in the utterance to the user device from which the utterance was received; and

updating, in the second storage, the user history of the user providing the utterance based on the utterance received from the user.

16. The article of manufacture of claim 15, wherein the process further comprises continually updating the user histories of each user in the second storage according to utterances from each user processed by the processing device and to update the personalization features for each user based on the updated user histories in the second storage.

17. The article of manufacture of claim 15, wherein personalization features of a same size have a same weight.

18. The article of manufacture of claim 15, wherein the personalization features include word based features including words from the user history.

19. The article of manufacture of claim 15, wherein the personalization features include class based features including classes of words from the user history.

20. The article of manufacture of claim 15, wherein the personalization features include features with both words and class information.