Reducing time for annotating speech data to develop a dialog application

• US 7,860,713 B2
• Filed: 07/01/2008
• Issued: 12/28/2010
• Est. Priority Date: 04/04/2003
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: AT&T Intellectual Property II LP (AT&T, Inc.)
• Inventors: Tur, Gokhan, Alonso, Tirso M., Hollister, Barbara B., Riccardi, Giuseppe, Wilson, James M., Bromberg, Ilana, Rahim, Mazin G., Stern, Daniel Leon, Hakkani-Tur, Dilek Z., Rose, Lawrence Lyon
• Primary Examiner(s): Lerner, Martin
• Application Number:
  US12/165,755
• Publication Number:
  US 20080270130A1
• Time in Patent Office: 910 Days
• Field of Search: 704/236, 704/238, 704/243, 704/244, 704/255, 704/257
• US Class Current: 704/236
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/19 Grammatical context, e.g. disambiguation of the recognition hypotheses based on word sequence rules
• CPC Class Codes: H04M 3/4936 Speech interaction details speech recognition per se G10L15/00

Abstract

Systems and methods for annotating speech data. The present invention reduces the time required to annotate speech data by selecting utterances for annotation that will be of greatest benefit. A selection module uses speech models, including speech recognition models and spoken language understanding models, to identify utterances that should be annotated based on criteria such as confidence scores generated by the models. These utterances are placed in an annotation list along with a type of annotation to be performed for the utterances and an order in which the annotation should proceed. The utterances in the annotation list can be annotated for speech recognition purposes, spoken language understanding purposes, labeling purposes, etc. The selection module can also select utterances for annotation based on previously annotated speech data and deficiencies in the various models.

Claims

1. In a system that uses annotated speech data, a method for annotating speech data by processing a portion of unannotated speech data with at least one model, the processing comprising:

evaluating a performance of at least one model with respect to each utterance in the portion of unannotated speech data using a criterion;

creating an annotation list that includes utterances that do not satisfy the criterion by;

using system deficiencies in combination with the criterion to identify utterances to be included in the annotation list; and

using previously annotated speech data in combination with the criterion or the system deficiencies to identify utterances to be included in the annotation list;

identifying an order in which the utterances on the annotation list are to be annotated;

generating via a processor a label for a particular utterance; and

including the particular utterance in the annotation list if the label does not match an existing label of the particular utterance.

2. The method of claim 1, further comprising:

searching the speech data for additional utterances that are similar to the utterances that do not satisfy the criterion; and

including the additional utterances in the annotation list.

3. The method of claim 1, wherein processing at least a portion of the speech data with the at least one model, wherein each utterance in the portion of the speech data is evaluated using a criterion, further comprises at least one of:

assigning each utterance in the portion of the speech data a confidence score, wherein utterances having a confidence score below a threshold confidence score are included in the annotation list;

evaluating a dialog context of each utterance in the portion of the speech data; and

evaluating a feature of each utterance in the portion of the speech data.

4. A system for annotating speech data by processing a portion of unannotated speech data with the at least one model, the system comprising:

a first module controlling a processor to evaluate a performance of the at least one model with respect to each utterance in the portion of unannotated speech data using a criterion;

a second module controlling the processor to create an annotation list that includes utterances that do not satisfy the criterion by;

using system deficiencies in combination with the criterion to identify utterances to be included in the annotation list; and

using previously annotated speech data in combination with the criterion or the system deficiencies to identify utterances to be included in the annotation list;

a third module controlling the processor to identify an order in which the utterances on the annotation list are to be annotated;

a fourth module controlling the processor to generate a label for a particular utterance; and

a fifth module controlling the processor to include the particular utterance in the annotation list if the label does not match an existing label of the particular utterance.

5. The system of claim 4, wherein the second module further uses system deficiencies in combination with the criterion to identify utterances to be included in the annotation list and uses previously annotated speech data in combination with the criterion or the system deficiencies to identify utterances to be included in the annotation list.

6. The system of claim 4, further comprising:

a sixth module controlling the processor to search the speech data for additional utterances that are similar to the utterances that do not satisfy the criterion; and

a seventh module controlling the processor to include the additional utterances in the annotation list.

7. The system of claim 4, wherein each utterance in the portion of the speech data is evaluated using a criterion further comprising at least one of:

assigning each utterance in the portion of the speech data a confidence score, wherein utterances having a confidence score below a threshold confidence score are included in the annotation list;

evaluating a dialog context of each utterance in the portion of the speech data; and

evaluating a feature of each utterance in the portion of the speech data.

8. A non-transitory computer-readable medium storing instructions for controlling a computing device to collect speech data to annotate speech data by processing a portion of unannotated speech data with the at least one model, the instructions comprising:

evaluating a performance of the at least one model with respect to each utterance in the portion of unannotated speech data using a criterion;

creating an annotation list that includes utterances that do not satisfy the criterion by;

using system deficiencies in combination with the criterion to identify utterances to be included in the annotation list;

using previously annotated speech data in combination with the criterion or the system deficiencies to identify utterances to be included in the annotation list;

identifying an order in which the utterances on the annotation list are to be annotated;

generating via a processor a label for a particular utterance; and

including the particular utterance in the annotation list if the label does not match an existing label of the particular utterance.

9. The non-transitory computer-readable medium of claim 8, the instructions further comprising:

searching the speech data for additional utterances that are similar to the utterances that do not satisfy the criterion; and

including the additional utterances in the annotation list.

10. The non-transitory computer-readable medium of claim 8, wherein each utterance in the portion of the speech data is evaluated using a criterion further comprises at least one of:

assigning each utterance in the portion of the speech data a confidence score, wherein utterances having a confidence score below a threshold confidence score are included in the annotation list;

evaluating a dialog context of each utterance in the portion of the speech data; and

evaluating a feature of each utterance in the portion of the speech data.