Method, device and system for noise-tolerant language understanding

• US 6,178,398 B1
• Filed: 11/18/1997
• Issued: 01/23/2001
• Est. Priority Date: 11/18/1997
• Status: Expired due to Term

• Current Assignee: Motorola Solutions, Inc.
• Original Assignee: Motorola, Inc. (Motorola Solutions, Inc.)
• Inventors: Karaali, Orhan, Peterson, Richard John, Bliss, Harry Martin, Russell, Dale William
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/972,515
• Time in Patent Office: 1,162 Days
• Field of Search: 704/251, 704/231, 704/232, 704/233, 704/258, 704/257, 704/200, 704/255
• US Class Current: 704/232
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding

Abstract

A method (900), device (200) and system (100) provide, in response to text/linguistic input, one of a set of pre-determined meanings which is the most likely intended meaning of that input. A trained meaning discriminator is generated from an annotated training corpus and a meaning discriminator trainer. The trained meaning discriminator generates a meaning vector from an input utterance. The intended meaning encoder analyzes the meaning vector to determine the most likely intended meaning and confidence measures.

Claims

1. A method for determining a meaning from an input utterance, comprising the steps of:

A) generating a trained meaning discriminator for correlating an input utterance to intended meaning using an annotated training corpus;

B) using the trained meaning discriminator to construct a meaning array from the input utterance; and

C) using the meaning array to provide information to a user as to a relative likelihood of each of a predetermined set of meanings.

2. The method of claim 1 wherein the input utterance is represented as text in the form of one of:

D) a string of words;

E) a word lattice; and

F) a combination of D and E.

3. The method of claim 1 where the input utterance is provided by an output of a speech recognizer in the form of:

D) a string of words;

E) a word lattice; and

F) a combination of D and E.

4. The method of claim 1 where the annotated training corpus contains data for each input utterance comprising:

a textual representation of the input utterance; and

at least one intended meaning of the input utterance.

5. The method of claim 4 wherein the annotated training corpus contains input utterances obtained from language in written form.

6. The method of claim 4 wherein the annotated training corpus contains utterances obtained from an output of a speech recognizer.

7. The method of claim 4 wherein the annotated training corpus contains at least one of a predetermined set of possible meanings associated with each utterance.

8. The method of claim 1 wherein generating the trained meaning discriminator comprises the steps of:

calculating a degree of correlation between a word of each utterance of the annotated training corpus and each of a plurality of possible meanings of the predetermined set of possible meanings; and

using the degree of correlation for each word of each utterance to construct a lexicon of words of the annotated training corpus, wherein each word of the lexicon is associated with indicators of a relative likelihood of each of the plurality of meanings of the predetermined set of possible meanings.

9. The method of claim 8 wherein calculating the degree of correlation between words of the annotated training corpus and possible meanings of the predetermined set of possible meanings is effected with a concordance scheme.

10. The method of claim 8 wherein calculating the degree of correlation between words of the annotated training corpus and the possible meanings is done with a statistical scheme.

11. The method of claim 8 wherein the meaning array is constructed utilizing the steps of:

obtaining for each word in the input utterance, indicators of the relative likelihood of each of the meanings of the predetermined set of possible meanings from the lexicon of words constructed from the annotated training corpus;

calculating indicators of the relative likelihood of each of the meanings of the predetermined set of possible meanings for the input utterance as an accumulation of the indicators of each word of the input utterance; and

analyzing the indicators of the relative likelihood of each of the meanings of the predetermined set of possible meanings for the input utterance using a meaning extractor.

12. The method of claim 1 wherein the meaning array is constructed utilizing the steps of:

extracting words from the input utterance to generate a vector of words;

processing the vector of words using a meaning discriminator; and

analyzing an output of the meaning discriminator using a meaning extractor.

13. The method of claim 12 wherein a primary component of the meaning discriminator is a neural network.

14. The method of claim 12 wherein a primary component of the meaning discriminator is a genetic algorithm unit.

15. The method of claim 12 wherein a primary component of the meaning discriminator is a decision tree unit.

16. The method of claim 12 wherein one of:

D) software implementing the method is embedded in a microprocessor;

E) software implementing the method is embedded in a digital signal processor;

F) the method is implemented by an application specific integrated circuit; and

G) the method is implemented by a combination of at least two of D-F.

17. The method of claim 1 further providing information to the user as to one of:

one best meaning according at least to a prespecified criterion; and

an ordered list of possible meanings.

18. The method of claim 1 further providing information to the user as to one of:

a uniqueness measure of a best meaning of the input utterance; and

a significance measure of the best meaning of the input utterance.

19. The method of claim 1 further providing information to the user in text form to be processed by a text to speech synthesizer.

20. The method of claim 1 further providing information to the user in text form to be processed by a dialogue manager.

21. The method of claim 1 wherein the trained meaning discriminator is trained using a method comprising the steps of:

extracting a set of words from the input utterance;

obtaining a meaning from the annotated training corpus; and

updating the trained meaning discriminator based on the set of words from the input utterance and the meaning from the annotated corpus.

22. The method of claim 21 where the set of words obtained by extracting are transformed into a vector whose length is equal to a number of vocabulary words in the annotated training corpus and whose elements have a value of 1 if a corresponding vocabulary word is in a sentence.

23. A device for determining a meaning from an input utterance, comprising:

A) a training subsystem for using an annotated training corpus to generate a trained meaning discriminator for correlating an input utterance to intended meanings to provide a trained statistical lexicon/trained neural network weights; and

B) a noise tolerant language understander, having a trained meaning discriminator arranged to receive the trained statistical lexicon/trained neural network weights, for using the trained statistical lexicon/trained neural network weights to construct a discrimination array and having a meaning extractor arranged to receive the discrimination array and output a meaning based on the discrimination array.

24. The device of claim 23 wherein the input utterance is represented as text in the form of one of:

C) a string of words;

D) a word lattice; and

E) a combination of C and D.

25. The device of claim 24 where the annotated training corpus contains data for each input utterance comprising:

a textual representation of the input utterance; and

at least one intended meaning of the input utterance.

26. The device of claim 25 wherein the annotated training corpus contains input utterances obtained from language in written form.

27. The device of claim 25 wherein the annotated training corpus contains utterances obtained from an output of a speech recognizer.

28. The device of claim 25 wherein the annotated training corpus contains at least one of a predetermined set of possible meanings associated with each utterance.

29. The device of claim 23 where the input utterance is provided by an output of a speech recognizer in the form of:

C) a string of words;

D) a word lattice; and

E) a combination of C and D.

30. The device of claim 23 wherein the trained meaning discriminator is generated by calculating a degree of correlation between each word of each utterance of the annotated training corpus and each of a plurality of possible meanings of the predetermined set of possible meanings;

and using a degree of correlation to construct a lexicon of the words of the annotated training corpus, wherein each word of the lexicon is associated with indicators of the relative likelihood of each of the plurality of meanings of the predetermined set of possible meanings.

31. The device of claim 30 wherein a concordance scheme is utilized to calculate the degree of correlation between words of the annotated training corpus and meanings of the predetermined set of possible meanings.

32. The device of claim 30 wherein a statistical scheme is utilized to calculate the degree of correlation between words of the annotated training corpus and the meanings of the predetermined set of possible meanings.

33. The device of claim 30 wherein construction of the meaning array includes:

obtaining for each word in the input utterance, the word'"'"'s indicators of the relative likelihood of each of the possible meanings of the predetermined set of possible meanings from the lexicon of words constructed from the annotated training corpus;

calculating indicators of the relative likelihood of each of the meanings of the predetermined set of possible meanings for the input utterance as an accumulation of the indicators of each word of the input utterance; and

analyzing the indicators of the relative likelihood of each of the meanings of the predetermined set of possible meanings for the input utterance using a meaning extractor.

34. The device of claim 23 wherein construction of the meaning array includes:

extracting words from the input utterance to generate a vector of words;

processing the vector of words using a meaning discriminator; and

analyzing an output of the meaning discriminator using a meaning extractor.

35. The device of claim 34 where a primary component of the meaning discriminator is a neural network.

36. The device of claim 34 where a primary component of the meaning discriminator is a genetic algorithm unit.

37. The device of claim 34 where a primary component of the meaning discriminator is a decision tree unit.

38. The device of claim 23 wherein one of:

C) device is implemented by software embedded in a microprocessor;

D) the device is implemented by software embedded in a digital signal processor;

E)the device is implemented by an application specific integrated circuit; and

F) the device is implemented by a combination of at least two of C-E.

39. The device of claim 23 wherein the relative likelihood generator further provides information to the user as to at least one of:

one best meaning according at least to a prespecified criterion; and

an ordered list of possible meanings.

40. The device of claim 23 wherein the relative likelihood generator further provides information to the user as to at least one of:

a uniqueness measure of a best meaning of the input utterance; and

a significance measure of the best meaning of the input utterance.

41. The device of claim 23 wherein the relative likelihood generator further provides information to the user in text form to be processed by a text to speech synthesizer.

42. The device of claim 23 wherein the relative likelihood generator further provides information to the user in text form to be processed by a dialogue manager.

43. The device of claim 23 wherein the trained meaning discriminator is trained using by extracting a set of words from the input utterance, obtaining a meaning from the annotated training corpus, and updating the trained meaning discriminator based on the set of words from the input utterance and the meaning from the annotated training corpus.

44. The device of claim 23 where the set of words obtained by extracting are transformed into a vector whose length is equal to a number of vocabulary words in the annotated training corpus and whose elements have a value of I if a corresponding vocabulary word is in a sentence.

45. A system having a device for determining a meaning from an input utterance, comprising:

A) a training subsystem for using an annotated training corpus to generate a trained meaning discriminator for correlating an input utterance to intended meanings to provide a trained statistical lexicon/trained neural network weights; and

B) a noise tolerant language understander, having a trained meaning discriminator arranged to receive the trained statistical lexicon/trained neural network weights, for using the trained statistical lexicon/trained neural network weights to construct a discrimination array and having a meaning extractor arranged to receive the discrimination array and output a meaning based on the discrimination array.

46. A device for determining a meaning from an input utterance, comprising:

A) a training subsystem for using an annotated training corpus to generate a trained meaning discriminator for correlating an input utterance to intended meanings to provide a trained statistical lexicon/trained neural network weights; and

B) a noise tolerant language understander, having a trained meaning discriminator arranged to receive the trained statistical lexicon/trained neural network weights, for using the trained statistical lexicon/trained neural network weights to construct a discrimination array and having a meaning extractor arranged to receive the discrimination array and output a meaning based on the discrimination array.

47. A system for determining a meaning from an input utterance, comprising:

A) a meaning discriminator generator for using an annotated training corpus to generate a trained meaning discriminator for correlating an input utterance to intended meanings;

B) a meaning array constructor, coupled to the trained meaning discriminator, for using the trained meaning discriminator to construct a meaning array from the input utterance; and

C) a relative likelihood generator, coupled to the meaning array constructor, for using the meaning array to provide information to a user as to a relative likelihood of each of a predetermined set of possible meanings.

48. A electrical communication unit for determining a meaning from an input utterance, comprising:

A) a training subsystem for using an annotated training corpus to generate a trained meaning discriminator for correlating an input utterance to intended meanings to provide a trained statistical lexicon/trained neural network weights; and

B) a noise tolerant language understander, having a trained meaning discriminator arranged to receive the trained statistical lexicon/trained neural network weights, for using the trained statistical lexicon/trained neural network weights to construct a discrimination array and having a meaning extractor arranged to receive the discrimination array and output a meaning based on the discrimination array.

49. The electrical communication unit of claim 48 wherein the electrical communication unit is one of:

a telephone, a computer, a personal digital assistant, a cellular phone, a mobile radio, a car navigation system, and a pager.