Updating population language models based on changes made by user clusters

• US 10,176,803 B2
• Filed: 06/05/2017
• Issued: 01/08/2019
• Est. Priority Date: 04/18/2013
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Bradford, Ethan R., Corston, Simon, Kay, David J., McCray, Donni, Trnka, Keith
• Primary Examiner(s): Shin, Seong Ah A
• Application Number:
  US15/613,939
• Publication Number:
  US 20170365253A1
• Time in Patent Office: 582 Days
• Field of Search: 704231, 704243, 704257, 704275
• CPC Class Codes: G10L 15/18 using natural language modelling

Abstract

Technology for improving the predictive accuracy of input word recognition on a device by dynamically updating the lexicon of recognized words based on the word choices made by similar users. The technology collects users'"'"' vocabulary choices (e.g., words that each user uses, or adds to or removes from a word recognition dictionary), associates users who make similar choices, aggregates related vocabulary choices, filters the words, and sends words identified as likely choices for that user to the user'"'"'s device. Clusters may include, for example, users in a particular location (e.g., sets of people who use words such as “Puyallup,” “Gloucester,” or “Waiheke”), users with a particular professional or hobby vocabulary, or application-specific vocabulary (e.g., word choices in map searches or email messages).

Claims

1. A method in a computing system of selectively generating or updating population language models used by language recognition systems, the method comprising:

receiving events representing user changes to local language models, wherein the events are received from devices of multiple users;

receiving information characterizing the multiple users;

using a hardware processor, identifying a user cluster from the received information;

identifying a subset of the events associated with the user cluster;

filtering the subset of events by excluding events associated with a blacklist of vocabulary not to be added;

identifying words associated with the subset of the events;

generating or updating a population language model for the user cluster based on at least some of the events and the associated words; and

providing the population language model or updates thereof to a computing device of one or more users in the user cluster for subsequently recognizing input information provided to the computing device by the one or more users in the user cluster.

2. The method of claim 1, wherein receiving events associated with local language models comprises receiving information about responses to word suggestions, addition of words to the local language model, or deletion of words from the local language model.

3. The method of claim 2, wherein information about word usage includes frequencies of use of individual words, word pairs (bigrams), triplets (trigrams), or higher-order n-grams.

4. The method of claim 2, wherein the events are received on a continuous, periodic, or irregular basis.

5. The method of claim 1, wherein the users in the cluster share a location.

6. The method of claim 1, wherein the users in the cluster share a common interest.

7. The method of claim 1, wherein identifying a cluster of users sharing similar characteristics further comprises identifying similar events associated with local language models.

8. The method of claim 7, wherein identifying similar events comprises determining word or n-gram associations.

9. The method of claim 1, further comprising:

identifying other words that are connected to the associated words,

wherein generating or updating the population language model includes determining the other words based on word-association for the words corresponding to the subset of the events or based on an inferential analysis of the corresponding words.

10. The method of claim 1, wherein generating modifications to a population language model associated with the cluster of users comprises adding words or n-grams to the population language model or removing words or n-grams from the population language model.

11. The method of claim 1, wherein generating modifications to a population language model associated with the cluster comprises modifying weighting of words or n-grams in the population language model.

12. The method of claim 1, wherein providing generated modifications to the computing device of one or more users in the cluster further comprises providing the associated population language model to a computing device of a user that is newly identified as having a connection to the cluster.

13. The method of claim 1, wherein providing generated modifications to a computing device of one or more users in the cluster comprises providing updates to a local population language model in a computing device of a user.

14. The method of claim 1, further comprising aggregating at least some of the events to identify modifications to the population model, including prioritizing events based on frequency of event.

20. The method of claim 1, wherein providing the generated modifications comprises prioritizing modifications by:

determining that a modification is of high importance; and

selectively providing the high-importance modification to the computing device of one or more users in the cluster.

15. A non-transitory computer-readable memory containing instructions that, when executed by a computing system, implement a method of selectively generating or updating local language models used to predictively complete user input, the method comprising:

receiving information about local language models including a first local language model and other local language models, wherein the information about local language models includes information about local language model events representing user changes to at least one of the local language models;

receiving information characterizing users associated with the user changes;

identifying a user cluster from the received characterization information;

identifying a set of the local language models including the first local language model and one or more other local language models, based at least in part on the other local language models having local language model events similar to first local language model events;

for each local language model in the set, identifying additional local language model events;

identifying words associated with the additional local language model events;

generating modifications to the first local language model using the additional local language model event information of one or more of the identified other local language models in the set, and also using the user cluster and the associated words;

filtering the generated modifications by excluding events associated with a blacklist of vocabulary not to be added before updating the first local language model with the generated modifications; and

updating the first local language model with the generated modifications based on providing the generated modifications to a computing device of one or more users in the user cluster.

16. The non-transitory computer-readable memory of claim 15, wherein receiving information about local language models comprises receiving a change log, local language models, or a combination thereof.

17. The non-transitory computer-readable memory of claim 15, further comprising:

identifying other words that are connected to the associated words,

wherein generating the modifications includes determining the other words based on word-association for the associated words corresponding to the local language model events or based on an inferential analysis of the associated words.

18. The non-transitory computer-readable memory of claim 15, wherein information about local language model events comprises information about addition of words to the local language model or deletion of words from the local language model.

19. The non-transitory computer-readable memory of claim 18, wherein information about word usage includes frequencies of individual words, word pairs (bigrams), triplets (trigrams), or higher-order n-grams.

21. The non-transitory computer-readable memory of claim 15, wherein updating the first local language model with the generated modifications comprises prioritizing modifications by:

determining that a modification is of high importance; and

selectively providing the high-importance modification to the first local language model.