Robust information extraction from utterances

• US 8,583,416 B2
• Filed: 12/27/2007
• Issued: 11/12/2013
• Est. Priority Date: 12/27/2007
• Status: Active Grant

• Current Assignee: Nant Holdings IP, LLC (NantWorks, LLC)
• Original Assignee: Fluential, LLC
• Inventors: Huang, Jun, Kim, Yookyung, Billawala, Youssef, Ehsani, Farzad, Master, Demitrios
• Primary Examiner(s): ALBERTALLI, BRIAN LOUIS
• Application Number:
  US11/965,711
• Publication Number:
  US 20090171662A1
• Time in Patent Office: 2,147 Days
• Field of Search: None
• US Class Current: 704/3
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding

Abstract

The performance of traditional speech recognition systems (as applied to information extraction or translation) decreases significantly with, larger domain size, scarce training data as well as under noisy environmental conditions. This invention mitigates these problems through the introduction of a novel predictive feature extraction method which combines linguistic and statistical information for representation of information embedded in a noisy source language. The predictive features are combined with text classifiers to map the noisy text to one of the semantically or functionally similar groups. The features used by the classifier can be syntactic, semantic, and statistical.

Claims

1. A method for classifying utterances, comprising the steps of:

providing access to a speech recognition engine;

receiving, by the speech recognition engine, an utterance as input;

using an automatic speech recognizer to covert the utterance into a text string;

generating an action-based concept tagged document of the text string;

classifying the text string into one or more semantic classes based at least in part on the action-based concept tagged document;

predicting paraphrased representations of the input with respect to a source language based on the one or more semantic classes; and

outputting a top candidate of the paraphrased representations in a target language.

2. The method of claim 1, wherein the classifying step includes using features sets and classification models to classify the text string into the classes.

3. The method of claim 2, further comprising extracting and combining statistical features, syntactic features, and semantic features from a training corpus to generate the feature sets.

4. The method of claim 1, wherein the classifying step further comprises extracting feature sets from the text string.

5. The method of claim 2, further comprising generating the classification models as a function of extracted features of a training corpus.

6. The method of claim 2, further comprising weighting features in the feature sets according to an information metric model from a training corpus.

7. The method of claim 1, wherein the classifying step further comprises classifying the text stings into one or semantic classes according to a hierarchy.

8. The method of claim 7, further comprising generating the hierarchy based on clustering of the semantic classes.

9. The method of claim 8 further comprising creating the clustering of the semantic classes based on misclassified data.

10. The method of claim 1, wherein the classifying step further comprising refining classification of the one or more semantic classes via rule-base template matching.

11. The method of claim 3, further comprising balancing the training corpus before generating the feature sets.

12. The method of claim 7, wherein the classifying step comprises performing a top-down classification of the one or more sematic classes according to the hierarchy.

13. The method of claim 7, wherein the classifying step comprises performing a bottom plus top-down classification of the one or more semantic classes according to the hierarchy.

14. The method of claim 1, wherein the classifying step comprises rejecting out-of-domain utterances based on a rejection model.

15. The method of claim 1, further comprising translating the classified text string based on the one or more semantic classes.

16. The method of claim 1, further comprising providing more than one n-best results from the speech recognition engine that are used as input text strings to the classifying step.

17. The method of claim 1, further comprising using a word recognition lattice from the speech recognition engine as input text strings to the classifying step.

18. The method of claim 1, further comprising using the semantic classes to control behavior of an IVR system.

19. The method of claim 1, wherein the receiving the utterance step includes detecting the utterance as noisy input and wherein the text string comprises noisy text input.