Learning from interactions for a spoken dialog system

• US 9,620,117 B1
• Filed: 06/27/2006
• Issued: 04/11/2017
• Est. Priority Date: 06/27/2006
• Status: Active Grant

• Current Assignee: AT&T Corporation (AT&T, Inc.), Rutgers University
• Original Assignee: AT&T Intellectual Property I LP (AT&T, Inc.)
• Inventors: Littman, Michael Lederman, Schapire, Robert E., Gilbert, Mazin, Levin, Esther
• Primary Examiner(s): Sanchez, Michael Ortiz
• Application Number:
  US11/426,748
• Time in Patent Office: 3,941 Days
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence

Abstract

In one embodiment, a semantic classifier input and a corresponding label attributed to the semantic classifier input may be obtained. A determination may be made whether the corresponding label is correct based on logged interaction data. An entry of an adaptation corpus may be generated based on a result of the determination. Operation of the semantic classifier may be adapted based on the adaptation corpus.

Claims

1. A method comprising:

obtaining, during an unsupervised classification phase of a spoken dialog system, a semantic classifier input and a corresponding label attributed to the semantic classifier input;

determining, via a processor, whether the corresponding label is correct based on logged interaction data, to yield a correctness result, wherein the logged interaction data comprises;

data describing user speech;

a non-speech user action indicating one of a negative training example and a positive training example; and

an input/output pair having an input and an output, the input comprising a speech recognition result in a lattice form and the output comprising one of an outcome of a call, a confirmation by a user, and a call hang-up, the output being a result of the input;

generating an entry for an adaptation corpus based on the correctness result; and

adapting operation of a semantic classifier based on the adaptation corpus.

2. The method of claim 1, wherein determining whether the corresponding label is correct indicates whether the label attributed by the semantic classifier is one of a correct label and an incorrect label.

3. The method of claim 1, wherein determining whether the corresponding label is correct further comprises:

calculating a confidence level indicative of an estimated probability that the label attributed by the semantic classifier is correct; and

determining that the label attributed by the semantic classifier is correct when the calculated confidence level is greater than a predetermined threshold.

4. The method of claim 1, wherein when the corresponding label is not correct, the method further comprises:

determining a correct label based on logged interaction data;

calculating a confidence level indicative of an estimated probability that the correct label is correct; and

recording the semantic classifier input, the correct label and the confidence level in a respective entry of the adaptation corpus.

5. The method of claim 1, wherein adapting operation of the semantic classifier based on the adaptation corpus further comprises:

using a boosting method to adapt the semantic classifier.

6. A system comprising:

a processor; and

a computer-readable storage medium having instructions stored which, when executed by the processor, cause the processor to perform operations comprising;

obtaining, during an unsupervised classification phase of a spoken dialog system, a semantic classifier input and a corresponding label attributed to the semantic classifier input;

determining, via a processor, whether the corresponding label is correct based on logged interaction data, to yield a correctness result, wherein the logged interaction data comprises;

data describing user speech;

a non-speech user action indicating one of a negative training example and a positive training example; and

an input/output pair having an input and an output, the input comprising a speech recognition result in a lattice form and the output comprising one of an outcome of a call, a confirmation by a user, and a call hang-up, the output being a result of the input;

generating an entry for an adaptation corpus based on the correctness result; and

adapting operation of a semantic classifier based on the adaptation corpus.

7. The system of claim 6, further comprising:

a dialog manager associated with the spoken dialog system configured to control the processor to receive the corresponding label and to determine one of a plurality of actions to perform based on the corresponding label.

8. The system of claim 7, wherein:

the dialog manager is further configured to control the processor to record interactions to the logged interaction data.

9. The spoken dialog system of claim 6, wherein determining whether the corresponding label is correct further comprises:

calculating a confidence level indicative of an estimated probability that the corresponding label is correct; and

determining that the corresponding label is correct when the calculated confidence level is greater than a predetermined threshold.

10. The system of claim 6, the computer-readable storage medium having additional instruction stored which result in the operations further comprising:

when the corresponding label is not correct;

determining a correct label based on the logged interaction data; and

calculating a confidence level indicative of an estimated probability that the correct label is correct; and

recording the semantic classifier input, the correct label and the confidence level in a respective entry of the adaptation corpus.

11. The system of claim 6, wherein the semantic classifier is arranged to adapt its operation based on the adaptation corpus by using a boosting method.

12. A computer-readable storage device having instructions stored which, when executed by a computing device, cause the processor to perform operations comprising:

obtaining, during an unsupervised classification phase of a spoken dialog system, a semantic classifier input and a corresponding label attributed to the input by a semantic classifier;

determining whether the corresponding label is correct based on logged interaction data, to yield a correctness result, wherein the logged interaction data comprises;

data describing user speech;

a non-speech user action indicating one of a negative training example and a positive training example; and

an input/output pair having an input and an output, the input comprising a speech recognition result in a lattice form and the output comprising one of an outcome of a call, a confirmation by a user, and a call hang-up, the output being a result of the input;

generating an entry for an adaptation corpus based on the correctness result result; and

adapting operation of a semantic classifier based on the adaptation corpus.

13. The computer-readable storage device of claim 12, the computer-readable storage device having additional instructions stored which result in the operations further comprising recording interactions of a component of the spoken dialog system to the logged interaction data.

14. The computer-readable storage device of claim 12, wherein determining whether the corresponding label is correct based on logged interaction data further comprises:

calculating a confidence level indicative of an estimated probability that the corresponding label attributed by the semantic classifier is correct; and

determining that the corresponding label attributed by the semantic classifier is correct when the calculated confidence level is greater than a predetermined threshold.

15. The computer-readable storage device of claim 12, wherein determining whether the corresponding label is correct based on logged interaction data, further comprises:

determining when the corresponding label is not correct;

determining a correct label based on the logged interaction data;

calculating a confidence level indicative of an estimated probability that the determined correct label is correct; and

recording the semantic classifier input, the correct label and the confidence level in a respective entry of the adaptation corpus.

16. The computer-readable storage device of claim 12, wherein determining whether the corresponding label is correct based on logged interaction data further comprises:

recording a confidence level to the adaptation corpus only when the corresponding label is determined to be incorrect.

17. The computer-readable storage device of claim 12, wherein adapting operation of the semantic classifier based on the adaptation corpus further comprises:

using a boosting method to adapt the semantic classifier.