SYSTEM AND METHOD FOR RESTRICTING LARGE LANGUAGE MODELS

• US 20110137653A1
• Filed: 12/04/2009
• Published: 06/09/2011
• Est. Priority Date: 12/04/2009
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: AT&T Intellectual Property I LP (AT&T, Inc.)
• Inventors: Gilbert, Mazin, LJOLJE, Andrej
• Granted Patent: US 8,589,163 B2
• Time in Patent Office: Days
• US Class Current: 704/255
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 2015/227 of the speaker; Human-factor methodology
• CPC Class Codes: G10L 2015/228 of application context

Abstract

Disclosed herein are systems, methods, and computer-readable storage media for performing speech recognition based on a masked language model. A system configured to practice the method receives a masked language model including a plurality of words, wherein a bit mask identifies whether each of the plurality of words is allowed or disallowed with regard to an adaptation subset, receives input speech, generates a speech recognition lattice based on the received input speech using the masked language model, removes from the generated lattice words identified as disallowed by the bit mask for the adaptation subset, and recognizes the received speech based on the lattice. Alternatively during the generation step, the system can only add words indicated as allowed by the bit mask. The bit mask can be separate from or incorporated as part of the masked language model. The system can dynamically update the adaptation subset and bit mask.

Claims

1. A computer-readable storage medium storing instructions which, when executed by a computing device, cause the computing device to generate a masked language model, the instructions comprising:

receiving a language model having a vocabulary of words;

identifying an adaptation subset of related words in the language model; and

adding a bit mask to each word in the large vocabulary of words based on the adaptation subset, thereby generating a masked language model.

2. The computer-readable storage medium of claim 1, the instructions further comprising:

identifying a plurality of adaptation subsets; and

adding a separate bit mask to each word in the large vocabulary of words corresponding to each of the plurality of adaptation subsets.

3. The computer-readable storage medium of claim 2, wherein the plurality of adaptation subsets is a multiple of a power of two.

4. The computer-readable storage medium of claim 1, wherein identifying an adaptation subset of related words is based on a geographical relationship.

5. The computer-readable storage medium of claim 1, wherein identifying an adaptation subset of related words is based on a speaker usage pattern.

6. The computer-readable storage medium of claim 1, wherein a speech recognition application allows only words in the language model having a matching bit mask value for a given adaptation subset.

7. The computer-readable storage medium of claim 1, wherein the bit mask comprises, for each word in the vocabulary, one of a 1 and a 0.

8. A system for recognizing speech using a masked language model, the system comprising:

a processor;

a module controlling the processor to receive a masked language model including a plurality of words, wherein a bit mask identifies whether each of the plurality of words is allowed or disallowed with regard to an adaptation subset;

a module controlling the processor to receive input speech;

a module controlling the processor to generate a speech recognition lattice based on the received input speech using the masked language model;

a module controlling the processor to remove from the generated lattice words identified as disallowed by the bit mask for the adaptation subset; and

a module controlling the processor to recognize the received speech based on the lattice.

9. The system of claim 8, wherein the bit mask is incorporated as part of the masked language model.

10. The system of claim 8, wherein the bit mask is separate from but associated with the masked language model.

11. The system of claim 8, the system further comprising a module controlling the processor to update the adaptation subset and bit mask based on a usage history.

12. The system of claim 8, wherein the adaptation subset and bit mask are associated with a user profile.

13. The system of claim 8, the system further comprising a module controlling the processor to enable and disable the bit mask.

14. The system of claim 8, wherein the bit mask comprises, for each of the plurality of words, one of a 1 and a 0.

15. A method for generating a masked language model, the method causing a computing device to perform steps comprising:

receiving a language model having a vocabulary of words;

identifying an adaptation subset of related words in the language model; and

generating a masked language model by adding a bit mask to the language model based on the adaptation subset.

16. The method of claim 15, further causing the computing device to perform steps comprising:

identifying a plurality of adaptation subsets; and

adding a separate bit mask to each word in the large vocabulary of words corresponding to each of the plurality of adaptation subsets.

17. The method of claim 15, further causing the computing device to dynamically generate the adaptation subset and associated bit mask.

18. The method of claim 15, wherein identifying the adaptation subset of related words is based on a geographical relationship.

19. The method of claim 15, wherein a speech recognition application allows only words in the language model having a matching bit mask value for a given adaptation subset.

20. The method of claim 15, wherein a speech recognition application allows only words in the language model having a matching bit mask value for each of a plurality of given adaptation subsets.