System and method of exploiting human-human data for spoken language understanding systems
A method is disclosed for generating labeled utterances from human-human utterances for use in training a semantic classification model for a spoken dialog system. The method comprises augmenting received human-human utterances with data that relates to call-type gaps in the human-human utterances, augmenting the received human-human utterances by placing at least one word in the human-human utterances that improves the training ability of the utterances according to the conversation patterns of the spoken dialog system, clausifying the human-human utterances, labeling the clausified and augmented human-human utterances and building the semantic classification model for the spoken dialog system using the labeled utterances.
1. A method for generating labeled utterances from human-human utterances for use in training a semantic classification model for a spoken dialog system, the method comprising:
identifying via a processor call-type gaps in human-human utterances to yield identified call-type gaps, wherein the identified call-type gaps include one of a missing call-type and an infrequent call-type;
augmenting the human-human utterances with data that relates to the identified call-type gaps in the human-human utterances;
augmenting the human-human utterances, to yield augmented human-human utterances by placing at least one word within the text of the human-human utterances that improves a training ability of the human-human utterances according to conversation patterns of the spoken dialog system;
clausifying the augmented human-human utterances to yield clausified, augmented human-human utterances;
labeling the clausified, augmented human-human utterances to yield labeled utterances; and
building a semantic classification model for the spoken dialog system using the labeled utterances.
2. The method of claim 1 , wherein clausifying the received human-human utterances occurs before the augmenting steps.
3. The method of claim 1 , wherein clausifying the received human-human utterances further comprises:
detecting sentence boundaries within a speech utterance text;
editing the speech utterance text to remove unneeded words; and
detecting conjunctions within the speech utterance text, wherein the clausifer outputs annotated text having identifiable clauses according to the sentence boundaries, edited text, and conjunctions within the speech utterance text.
4. The method of claim 1 , wherein the data that relates to call-type gaps used to augment the received human-human utterances is borrowed from other spoken dialog system applications.
5. The method of claim 4 , wherein the other spoken dialog system applications have a related function to the spoken dialog system.
6. The method of claim 1 , wherein only a portion of the clausified human-human utterances are labeled and used to build the semantic classification model.
7. The method of claim 6 , wherein the portion of the clausified human-human utterances that are labeled and used to build the semantic classification model relate to identifying the intent of the speaker.
8. The method of claim 1 , wherein the at least one word is either no or yes.
9. The method of claim 1 , wherein the at least one word relates to a phrase related to a computer-human interaction.
10. The method of claim 1 , wherein missing or infrequent call-types include call types which are missing or infrequent in the human-human utterances, but would be more common in human-machine dialogs.