Robust Information Extraction from Utterances

• US 20090171662A1
• Filed: 12/27/2007
• Published: 07/02/2009
• Est. Priority Date: 12/27/2007
• Status: Active Grant

• Current Assignee: Nant Holdings IP, LLC (NantWorks, LLC)
• Original Assignee: Sehda, Inc.
• Inventors: Huang, Jun, Ehsani, Farzad, Master, Demitrios, Billawala, Youssef, Kim, Yookyung
• Granted Patent: US 8,583,416 B2
• Time in Patent Office: Days
• US Class Current: 704/251
• 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:

receiving a text string;

classifying the text string into one or more semantic classes.

1.1. 1 The method of claim 1 wherein the text string is generated from a speech recognition engine with the utterances as the input.

2. The method of claim 1, wherein in the classifying step, feature sets and classification models are used in the classifying process.

3. The method of claim 2, wherein statistical features, syntactic features, and semantic features of a training corpus are extracted and combined to generate features sets.

4. The method of claim 1, wherein in the classifying step, feature sets of the text strings are extracted.

5. The method of claim 2, wherein classification models are generated as a function of the extracted features of a training corpus.

6. The method of claim 2, wherein features are weighed according to information metric model estimated from a training corpus.

7. The method of claim 1, wherein in the classifying step, classification is done hierarchically.

8. The method of claim 7, wherein the hierarchy is generated based on clustering of the semantic classes.

9. The method of claim 8, wherein the clustering of the semantic classes is created based on misclassified data.

10. The method of claim 1, wherein in the classifying step, classification is further refined with rule-based templates matching.

11. The method of claim 3, wherein in the training steps, the training data is balanced before generating features sets.

12. The method of claim 9, wherein the hierarchical classification is performed on a top-down basis.

13. The method of claim 9, wherein the hierarchical classification is performed on a bottom plus top-down basis.

14. The method of claim 1, wherein a rejection model is used in the classifying step to reject out-of-domain utterances.

15. The method of claim 1, the semantic classes are used to translate the classified text string.

16. The method of claim 1.1, wherein there are 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 1.1, wherein a word recognition lattice from the speech recognition engine is used as input text strings to the classifying step.

18. The method of claim 1, the semantic classes are used to control the behavior of an IVR system.