System and method for isolating uncertainty between speech recognition and natural language processing

• US 8,560,311 B2
• Filed: 09/23/2010
• Issued: 10/15/2013
• Est. Priority Date: 09/23/2009
• Status: Active Grant

• Current Assignee: Arlington Technologies, LLC (Dominion Harbor Enterprises, LLC)
• Original Assignee: Avaya Incorporated
• Inventors: Williams, Robert W., Keane, John E.
• Primary Examiner(s): Lerner, Martin
• Application Number:
  US12/889,172
• Publication Number:
  US 20110213616A1
• Time in Patent Office: 1,118 Days
• Field of Search: 704/9, 704/231, 704/236, 704/252, 704/257, 704/232, 706/20
• US Class Current: 704/232
• CPC Class Codes: G10L 15/18 using natural language modelling

Abstract

A speech recognition system includes a natural language processing component and an automated speech recognition component distinct from each other such that uncertainty in speech recognition is isolated from uncertainty in natural language understanding, wherein the natural language processing component and an automated speech recognition component communicate corresponding weighted meta-information representative of the uncertainty.

Claims

1. A speech recognition system comprising at least one computer including at least one processor, the at least one computer comprising a natural language processing component, a machine learning engine, and an automated speech recognition component, the natural language processing component and the automated speech recognition component being distinct from each other such that uncertainty in speech recognition is isolated from uncertainty in natural language understanding, wherein the natural language processing component and the automated speech recognition component communicate corresponding weighted meta-information representative of the uncertainty, the corresponding weighted meta-information generated at least partially by the machine learning engine, wherein the machine learning engine is configured to:

receive, from the automated speech recognition component, weighted meta-information;

adjust a weighting of the weighted meta-information based at least partially on a statistical learning algorithm; and

communicate the adjusted weighted meta-information to the natural language processing component.

2. The speech recognition system of claim 1, wherein the statistical learning algorithm is developed by the machine learning engine by analyzing a bi-direction flow of the corresponding weighted meta-information between the natural language processing component and the automated speech recognition component.

3. The speech recognition system of claim 2, wherein the statistical weighting algorithm comprises at least one of a Bayes-based statistical model and a neural network.

4. The speech recognition system of claim 1, wherein the corresponding weighted meta-information generated by the natural language processing component comprises at least one confidence metric.

5. The speech recognition system of claim 1, wherein the machine learning engine is further configured to receive, from the automated speech recognition component, a plurality of words or word segments from the automated speech recognition component, and wherein the weighted meta-information received from the automated speech recognition component comprises match-likelihood values for at least a portion of the plurality of words or word segments.

6. The speech recognition system of claim 1, wherein the natural language processing component is further configured to communicate, to a speech engine, a response generated at least partially based on the corresponding weighted meta-information.

7. A system for isolating uncertainty between speech recognition and natural language processing, the system comprising at least one computer including at least one processor, the at least one computer comprising:

(a) an automated speech recognition component configured to generate, based on speech input, at least one of words and word sequences, the at least one of words and word sequences associated with meta-information;

(b) a natural language processing component distinct from the automated speech recognition component, the natural language processing component configured to process the at least one of words and word sequences based at least partially on weighted meta-information; and

(c) a machine learning engine configured to;

(i) receive the meta-information from the automated speech recognition system;

(ii) generate the weighted meta-information for the at least one of words and word sequences based at least partially on a learned weighting algorithm and the at least one of words and word sequences, wherein the learned weighting algorithm is influenced by the machine learning engine by analyzing meta-information received from the automated speech recognition component and the natural language processing component.

8. The system of claim 7, wherein the machine learning engine is further configured to:

(iii) receive, from the natural language processing component, corresponding meta-information; and

(iv) influence the learned weighting algorithm based at least partially on the meta-information received from the automated speech recognition component and the corresponding meta-information received from the natural language processing component.

9. The system of claim 8, wherein the corresponding meta-information received from the natural language processing component comprises at least one confidence metric.

10. The system of claim 7, wherein the weighted meta-information comprises match probabilities for the at least one of words and word sequences.

11. The system of claim 10, wherein the natural language processing component is further configured to process the at least one of words and word sequences based at least partially on the match probabilities.

12. The system of claim 11, wherein the natural language processing component is further configured to communicate, to a speech engine, a response generated by the processing of the at least one of words and word sequences.

13. The system of claim 7, wherein the natural language processing component is further configured to pre-process the at least one of words and word sequences by at least one of the following:

spelling correction, alias substitution, removal of words determined to be informationally unimportant, word reordering, or any combination thereof.

14. The system of claim 7, wherein the learned weighting algorithm comprises at least one of a Bayes-based statistical model and a neural network.

15. A computer-implemented method for isolating uncertainty between speech recognition and natural language processing, the method using a speech recognition engine and a natural language processing engine that are distinct from one another, the method comprising:

receiving speech input from at least one user;

generating, with the automated speech recognition engine and at least one processor, at least one word or word sequence associated with meta-information based at least partially on the speech input, the meta-information comprising at least one match probability for each of the at least one word or word sequence;

generating, with a machine learning engine, weighted meta-information based at least partially on the at least one word or word sequence and the meta-information;

generating, with the natural language processing engine and at least one processor, corresponding meta-information based at least partially on the weighted meta-data and the at least one word or word Sequence;

communicating the corresponding meta-information from the natural language processing engine to the machine learning engine; and

generating, with the natural language processing engine, an output response based at least partially on the weighted meta-information.

16. The computer-implemented method of claim 15, wherein the corresponding meta-information comprises a confidence metric generated by the natural language processing engine.

17. The computer-implemented method of claim 15, wherein the learned weighting algorithm is influenced by at least the weighted meta-information and the corresponding meta-information.

18. The computer-implemented method of claim 15, wherein the weighted meta-information comprises match probabilities for the at least one word or word sequence sequences.

19. The computer-implemented method of claim 15, wherein the natural language processing component is configured to communicate, to a speech engine, the output response.