System and method of semi-supervised learning for spoken language understanding using semantic role labeling
View Patent ↗A system and method are disclosed for providing semi-supervised learning for a spoken language understanding module using semantic role labeling. The method embodiment relates to a method of generating a spoken language understanding module. Steps in the method comprise selecting at least one predicate/argument pair as an intent from a set of the most frequent predicate/argument pairs for a domain, labeling training data using mapping rules associated with the selected at least one predicate/argument pair, training a call-type classification model using the labeled training data, re-labeling the training data using the call-type classification model and iteratively several of the above steps until training set labels converge.
1. A method of generating a spoken language understanding module, the method comprising:
a. selecting, via a processor, at least one predicate/argument pair as an intent from a set of a most frequent predicate/argument pairs for a domain associated with a spoken dialog system in a first language;
b. labeling training data using mapping rules for the first language associated with the selected at least one predicate/argument pair and that specify rules for selecting a call-type label for an utterance;
c. training a call-type classification model using the labeled training data;
d. re-labeling the training data using the call-type classification model; and
e. iteratively processing steps (c) and (d) until training set labels converge.
2. The method of claim 1 , further comprising assigning the verbs be and have as special predicates.
3. The method of claim 2 , wherein assigning the verbs distinguishes the verbs from utterances which do not have a predicate.
4. The method of claim 1 , further comprising capturing infrequent call types using an active-learning approach.
5. The method of claim 1 , wherein the method is semi-supervised.
6. The method of claim 1 , wherein selecting at least one predicate/argument pair further comprises generating mapping rules and for labeling the training data.
7. The method of claim 1 , wherein selecting the at least one predicate/argument pair is processed independent of the domain.
8. A tangible computer-readable medium storing instructions for controlling a computing device to generate a spoken language understanding module, the instructions comprising:
a. selecting at least one predicate/argument pair as an intent from a set of the most frequent predicate/argument pairs for a domain associated with a spoken dialog system in a first language;
b. labeling training data using mapping rules for the first language associated with the selected at least one predicate/argument pair and that specify rules for selecting a call-type label for an utterance;
c. training a call-type classification model using the labeled training data;
d. re-labeling the training data using the call-type classification model; and
e. iteratively processing steps (c) and (d) until training set labels converge.
9. The tangible computer-readable medium of claim 8 , the instructions further comprising assigning the verbs be and have as special predicates.
10. The tangible computer-readable medium of claim 9 , wherein assigning the verbs distinguishes the verbs from utterances which do not have a predicate.
11. The tangible computer-readable medium of claim 8 , the instructions further comprising capturing infrequent call types using an active-learning approach.
12. The tangible computer-readable medium of claim 8 , wherein the instructions are semi-supervised.
13. The tangible computer-readable medium of claim 8 , wherein selecting at least one predicate/argument pair further comprises generating mapping rules and for labeling the training data.
14. The tangible computer-readable medium of claim 8 , wherein selecting the at least one predicate/argument pair is processed independent of the domain.
15. A system for generating a spoken language understanding module, the system comprising:
a. a first module configured to select at least one predicate/argument pair as an intent from a set of the most frequent predicate/argument pairs for a domain associated with a spoken dialog system in a first language;
b. a second module configured to label training data using mapping rules for a first language associated with the selected at least one predicate/argument pair and that specify rules for selecting a call-type label for an utterance;
c. a third module configured to train a call-type classification model using the labeled training data;
d. a fourth module configured to re-label the training data using the call-type classification model; and
e. a fifth module configured to iteratively process steps (c) and (d) until training set labels converge.
16. The system of claim 15 , further comprising a sixth module configured to assign the verbs be and have as special predicates.
17. The system of claim 16 , wherein the sixth module further distinguishes the verbs from utterances which do not have a predicate.
18. The system of claim 15 , further comprising a sixth module configured to capture infrequent call types using an active-learning approach.
19. The system of claim 15 , wherein the modules perform their processes in a semi-supervised manner.
20. The system of claim 15 , wherein the first module further generates mapping rules and for labeling the training data.
21. The system of claim 15 , wherein the first module performs its process independent of the domain.