Active labeling for spoken language understanding
View Patent ↗An active labeling process is provided that aims to minimize the number of utterances to be checked again by automatically selecting the ones that are likely to be erroneous or inconsistent with the previously labeled examples. In one embodiment, the errors and inconsistencies are identified based on the confidences obtained from a previously trained classifier model. In a second embodiment, the errors and inconsistencies are identified based on an unsupervised learning process. In both embodiments, the active labeling process is not dependent upon the particular classifier model.
1. In a system that includes an automatic speech recognition component that extracts words from user speech, a classification method comprising:
selecting a first set of candidate utterances, each of said first set of candidate utterances labeled with at least one previously assigned classification type;
generating, based on a trained classifier, a confidence score for a plurality of classification types for each of said first set of candidate utterances; and
identifying a second set of candidate utterances from said first set of candidate utterances as having a potential classification error, wherein said identifying is based on an analysis of previously assigned classification types and generated confidence scores.
2. The classification method of claim 1 , wherein said confidence score is generated by said trained classifier.
3. The classification method of claim 1 , wherein said confidence score is calculated based on an output of said trained classifier.
4. The classification method of claim 1 , wherein said identifying comprises identifying those candidate utterances where a Kullback-Leibler divergence is greater than a threshold.
5. The classification method of claim 4 , wherein said identifying is based on the computation:
KL
(
P
Q
)
=
∑
i
∈
L
p
i
×
log
(
p
i
q
i
)
+
(
1
-
p
i
)
×
log
(
1
-
p
i
1
-
q
i
)
where L is the set of all classification types, q i is the probability of the i th classification type obtained from said trained classifier, and p=1 if that classification type is previously labeled and p=0 if otherwise.
6. The classification method of claim 1 , wherein said identifying comprises identifying those candidate utterances where a top choice of said trained classifier is not among the at least one previously assigned classification type.
7. The classification method of claim 1 , wherein said first set of candidate utterances includes labeled and unchecked data.
8. The classification method of claim 1 , wherein said classification types include call classification types.
9. In a system that includes a speech classification component that extracts words from use speech, a classification method, comprising:
selecting a first set of candidate utterances, each of said first set of candidate utterances labeled with at least one previously assigned classification type;
training a classifier using said first set of candidate utterances to produce a trained classifier;
classifying a second set of candidate utterances using said trained classifier; and
identifying a second set of candidate utterances from said first set of candidate utterances as having a potential classification error, wherein said identifying is based on an analysis of the previously assigned classification types and the results of said classifying.
10. The classification method of claim 9 , wherein said first set of candidate utterances includes labeled and unchecked data.
11. The classification method of claim 9 , wherein said classifier is an iterative classifier.
12. The classification method of claim 9 , further comprising generating, based on a trained classifier, a confidence score for a plurality of classification types for each of said first set of candidate utterances.
13. The classification method of claim 9 , wherein said classification types include call classification types.
14. The classification method, comprising:
classifying a set of candidate utterances using a classifier, said set of candidate utterances including labeled and unchecked data; and
automatically selecting a subset of said set of candidate utterances as likely including erroneous or inconsistent classifications, said automatic selection being based on an analysis of an output of said classifier.
15. The classification method of claim 14 , wherein said classifier is an iterative classifier.
16. The classification method of claim 14 , wherein said output of said classifier includes confidence scores.
17. The classification method of claim 14 , wherein said output of said classifier includes information used to derive confidence scores.
18. The classification method of claim 14 , wherein said output of said classifier includes classification labels.
19. A computer-readable medium that stores a program for controlling a computer device to perform the following classification method in a system that includes an automatic speech recognition component that extracts words from user speech, the method comprising:
selecting a first set of candidate utterances, each of said first set of candidate utterances labeled with at least one previously assigned classification type;
generating, based on a trained classifier, a confidence score for a plurality of classification types for each of said first set of candidate utterances; and
identifying a second set of candidate utterances from said first set of candidate utterances as having a potential classification error, wherein said identifying is based on an analysis of previously assigned classification types and generated confidence scores.
20. A computer-readable medium that stores a program for controlling a computer device to perform the following classification method in a system that includes a speech classification component that extracts words from user speech, the method comprising:
selecting a first set of candidate utterances, each of said first set of candidate utterances labeled with at least one previously assigned classification type;
training a classifier using said first set of candidate utterances to produce a trained classifier;
classifying a second set of candidate utterances using said trained classifier; and
identifying a second set of candidate utterances from said first set of candidate utterances as having a potential classification error, wherein said identifying is based on an analysis of the previously assigned classification types and the results of said classifying.
21. A computer-readable medium that stores a program for controlling a computer device to perform the following classification method, the method comprising:
classifying a set of candidate utterances using a classifier, said set of candidate utterances including labeled and unchecked data; and
automatically selecting a subset of said set of candidate utterances as likely including erroneous or inconsistent classifications, said automatic selection being based on an analysis of an output of said classifier.