Addressing Missing Features in Models

• US 20160267904A1
• Filed: 04/08/2015
• Published: 09/15/2016
• Est. Priority Date: 03/13/2015
• Status: Active Grant

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: Google Inc. (Alphabet Inc.)
• Inventors: Biadsy, Fadi, Caseiro, Diamantino Antonio
• Granted Patent: US 9,805,713 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence
• CPC Class Codes: G10L 15/30 Distributed recognition, e.g. in client-server systems, for mobile phones or network applications

Abstract

Systems and methods for addressing missing features in models are provided. In some implementations, a model configured to indicate likelihoods of different outcomes is accessed. The model includes a respective score for each of a plurality of features, and each feature corresponds to an outcome in an associated context. It is determined that the model does not include a score for a feature corresponding to a potential outcome in a particular context. A score is determined for the potential outcome in the particular context based on the scores for one or more features in the model that correspond to different outcomes in the particular context. The model and the score are used to determine a likelihood of occurrence of the potential outcome.

Claims

1. A method performed by one or more computers, comprising:

receiving data indicating a candidate transcription for an utterance and a context for the utterance;

accessing a language model that includes a respective score for each of a plurality of features, each feature corresponding to a word or phrase occurring in an associated context that includes one or more preceding words;

determining that the language model does not include a score for a feature corresponding to the candidate transcription in the particular context;

determining a score corresponding to the candidate transcription in the particular context, wherein the score is determined based on one or more scores included in the language model for one or more of the plurality of features that are associated with the particular context;

determining, using the language model and the determined score, a probability score indicating a likelihood of occurrence of the candidate transcription in the particular context;

selecting, based on the probability score, a transcription for the utterance from among a plurality of candidate transcriptions; and

providing the selected transcription to a client device.

2. The method of claim 1, wherein the language model is a log-linear model.

3. The method of claim 1, wherein determining the score corresponding to the candidate transcription in the particular context comprises:

identifying features in the language model that correspond to different words or phrases occurring in the particular context; and

accessing scores in the language model for the identified features.

4. The method of claim 3, wherein the identified features constitute an exhaustive set of features of the model that are associated with the particular context.

5. The method of claim 3, wherein determining the score corresponding to the candidate transcription in the particular context further comprises:

identifying the minimum score from among the accessed scores for the identified features; and

determining the score corresponding to the candidate transcription in the particular context based on the identified minimum score.

6. The method of claim 5, wherein determining the score corresponding to the candidate transcription in the particular context comprises determining the score by subtracting one or more predefined values from the minimum score.

7. The method of claim 3, wherein (i) identifying the features in the language model that correspond to different words or phrases occurring in the particular context and (ii) accessing the scores for the identified features are performed in response to determining that the language model does not include a score for a feature corresponding to the candidate transcription in the particular context.

8. The method of claim 3, wherein (i) identifying the features in the language model that correspond to different words or phrases occurring in the particular context and (ii) accessing the scores for the identified features are performed prior to receiving the candidate transcription for the utterance.

9. The method of claim 1, wherein determining the score corresponding to the candidate transcription in the particular context comprises:

accessing a stored score that is assigned to the particular context, the stored score being assigned to the particular context prior to the utterance being spoken; and

using the stored score as the score corresponding to the candidate transcription;

10. The method of claim 1, wherein determining the score corresponding to the candidate transcription in the particular context comprises determining a score indicating a likelihood of occurrence that is less than or is equal to the lowest likelihood of occurrence indicated by scores in the language model that are assigned to features that correspond to the particular context.

11. The method of claim 1, wherein the language model is a log-linear model, and wherein accessed scores for the identified features are weights of the log-linear model that are associated with the identified features.

12. The method of claim 1, wherein the language model has been trained to indicate a likelihood of a word or phrase occurring in a language sequence based at least in part on one or more prior words in the language sequence.

13. The method of claim 1, wherein each of the plurality of features corresponds to a respective language sequence that occurs in training data that was used to train the language model;

and

wherein the candidate transcription and the particular context form a language sequence that was not included in the training data that was used to train the model.

14. The method of claim 1, further comprising:

pre-computing a minimum score for each of multiple contexts while training the language model; and

normalizing scores for features in the language model using the pre-computed minimum scores.

15. A system comprising:

one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising;

receiving data indicating a candidate transcription for an utterance and a context for the utterance;

accessing a language model that includes a respective score for each of a plurality of features, each feature corresponding to a word or phrase occurring in an associated context that includes one or more preceding words;

determining that the language model does not include a score for a feature corresponding to the candidate transcription in the particular context;

determining a score corresponding to the candidate transcription in the particular context, wherein the score is determined based on one or more scores included in the language model for one or more of the plurality of features that are associated with the particular context;

determining, using the language model and the determined score, a probability score indicating a likelihood of occurrence of the candidate transcription in the particular context;

selecting, based on the probability score, a transcription for the utterance from among a plurality of candidate transcriptions; and

providing the selected transcription to a client device.

16. The system of claim 15, wherein the language model is a log-linear model.

17. The system of claim 16, wherein determining the score corresponding to the candidate transcription in the particular context comprises determining a score indicating a likelihood of occurrence that is less than or is equal to the lowest likelihood of occurrence indicated by scores in the language model that are assigned to features that correspond to the particular context.

18. A non-transitory computer readable storage medium storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

receiving data indicating a candidate transcription for an utterance and a context for the utterance;

accessing a language model that includes a respective score for each of a plurality of features, each feature corresponding to a word or phrase occurring in an associated context that includes one or more preceding words;

determining that the language model does not include a score for a feature corresponding to the candidate transcription in the particular context;

determining a score corresponding to the candidate transcription in the particular context, wherein the score is determined based on one or more scores included in the language model for one or more of the plurality of features that are associated with the particular context;

determining, using the language model and the determined score, a probability score indicating a likelihood of occurrence of the candidate transcription in the particular context;

selecting, based on the probability score, a transcription for the utterance from among a plurality of candidate transcriptions; and

providing the selected transcription to a client device.

19. The non-transitory computer readable storage medium of claim 18, wherein the language model is a log-linear model.

20. The non-transitory computer readable storage medium of claim 19, wherein determining the score corresponding to the candidate transcription in the particular context comprises determining a score indicating a likelihood of occurrence that is less than or is equal to the lowest likelihood of occurrence indicated by scores in the language model that are assigned to features that correspond to the particular context.