Systems and methods for word recognition

• US 5,680,511 A
• Filed: 06/07/1995
• Issued: 10/21/1997
• Est. Priority Date: 06/07/1995
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation), Express Communications Supply, Inc. (Trip Technologies, Inc.)
• Original Assignee: Dragon Systems, Inc. (Microsoft Corporation)
• Inventors: Baker, Janet M., Yamron, Jonathan P., Baker, James K., Gillick, Laurence S.
• Primary Examiner(s): Zele, Krista M.
• Assistant Examiner(s): Hunter, Daniel S.
• Application Number:
  US08/477,287
• Time in Patent Office: 867 Days
• Field of Search: 364/419.02, 364/419.03, 364/419.08, 395/2.6, 395/2.61, 395/2.64, 395/2.66, 395/2.4, 395/2.44, 395/2.79, 395/2.69, 395/2.86, 395/2.85, 395/2.52, 395/2.84
• US Class Current: 704/257
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning

Abstract

In one aspect, the invention provides word recognition systems that operate to recognize an unrecognized or ambiguous word that occurs within a passage of words. The system can offer several words as choice words for inserting into the passage to replace the unrecognized word. The system can select the best choice word by using the choice word to extract from a reference source, sample passages of text that relate to the choice word. For example, the system can select the dictionary passage that defines the choice word. The system then compares the selected passage to the current passage, and generates a score that indicates the likelihood that the choice word would occur within that passage of text. The system can select the choice word with the best score to substitute into the passage. The passage of words being analyzed can be any word sequence including an utterance, a portion of handwritten text, a portion of typewritten text or other such sequence of words, numbers and characters. Alternative embodiments of the present invention are disclosed which function to retrieve documents from a library as a function of context.

Claims

1. A method for ascertaining an ambiguous word that occurs in a data string which includes at least one context word, comprising the steps of

generating a list of one or more choice words to represent choices for the ambiguous word,

providing a reference source having one or more passages, each passage including a relevance indicator and a series of words associated with the relevance indicator,

selecting as a function of one of said choice words and said relevance indicators, at least one of said passages from said reference source,

generating one or more correlation signals, as a function of said one or more selected passages, and at least one of said context words, to represent the likelihood that said choice word is a match for the ambiguous word, and

selecting, as a function of said one or more correlation signals, at least one of said choice words to represent the ambiguous word being identified.

2. A method according to claim 1 for use with a continuous speech recognition system, wherein

said step of generating said list of choice words, includes the steps of comparing select characteristics of a portion of speech to one or more word models each being representative of a word in a vocabulary, and

selecting as a function of said comparison a subset of said words from said vocabulary.

3. The method according to claim 1 wherein said step of generating said correlation signals includes the further step of

generating a list of passage words representative of one or more of the content words in said selected passages.

4. A method according to claim 1 wherein said step of providing a reference source includes the step of

providing a reference source comprising a dictionary.

5. A method according to claim 1 wherein said step of providing a reference source includes the step of

providing a reference source comprising an encyclopedia.

6. A method according to claim 1 wherein said step of selecting passages includes the step of

selecting a choice word in a first language,

translating said selected choice word into a second language,

accessing a reference source comprising information recorded in the second language and selecting said passages therefrom, and

translating said selected passages into the first language.

7. A method according to claim 1 wherein the step of selecting passages from a reference source includes the step of

comparing the choice word with the relevance indicators, and selecting passages having a relevance indicator that contains a word that substantially matches the choice word.

8. A method according to claim 1 wherein the step of generating said correlation signals includes the step of

determining, for each choice word, a word count representative of the number of distinct words in the associated passages that match with one of the context words in the data string and generating said correlation signal as a function of said word count.

9. A method according to claim 8 wherein said step of generating said correlation signals includes the further step of

determining a frequency count for each passage word representative of the number of times the passage word occurs in a passage, and

generating said correlation signals as function of said word count weighted by said frequency count.

10. A method according to claim 1 wherein said step of generating correlation signals includes the further steps of

generating an occurrence signal for each of said choice words as a function of an n-gram language model and one or more of said context words to represent the likelihood that said choice word would be selected from a set of vocabulary words, and

combining said occurrence signal with said correlation signal, representative of the likelihood that said choice word is a match for the ambiguous word, to adjust said correlation signal as a function of said likelihood of occurrence for said choice word.

11. A method according to claim 1 wherein said step of generating said correlation signals includes the step of

adjusting each of said correlation signals as a function of a word count signal representative of the number of words separating a matching context word from the ambiguous word being recognized.

12. A method according to claim 1 wherein the step of generating the correlation signals includes the step of comparing the selected passages with the context words to detect a matching n-gram.

13. A method for ascertaining an ambiguous word that occurs in a data string that includes one or more context words, comprising the steps of

generating a list of one or more choice words to represent choices for the ambiguous word,

providing a reference source having one or more passages each of which includes a relevance indicator and a series of passage words associated with the relevance indicator,

selecting as a function of each of said relevance indicators and said context words, at least one passage from said reference source for at least one of said context words,

generating a correlation signal for each of said choice words as a function of said passages and said choice words, to represent the likelihood that a respective choice word is a match for the ambiguous word, and

selecting as a function of said correlation signals, at least one of said choice words to represent said ambiguous word.

14. A method according to claim 13 for use with a continuous speech recognition system, wherein

said step of generating said list of choice words, includes the steps of comparing select characteristics of a portion of speech to one or more word models each being representative of a word in a vocabulary, and

selecting as a function of said comparison a subset of said words from said vocabulary.

15. The method according to claim 13 wherein said step of generating said correlation signal includes the further step of

generating a list of passage words representative of one or more of the content words in said passage.

16. A method according to claim 13 wherein said step of selecting one or more passages includes the step of

providing a reference source comprising a dictionary.

17. A method according to claim 13 wherein said step of selecting one or more passages includes the step of

providing a reference source comprising an encyclopedia.

18. A method according to claim 13 wherein said step of selecting one or more passages includes the steps of

selecting a context word in a first language,

translating said selected context word into a second language,

accessing a reference source comprising information recorded in the second language and selecting said one or more passages therefrom, and

translating said selected passages into the first language.

19. A method according to claim 13 wherein the step of selecting passages from a reference source includes the step of

comparing the context word with each of the relevance indicators, and selecting passages having a relevance indicator that contains a word that substantially matches the context word.

20. A method according to claim 13 wherein the step of generating said correlation signals includes the step of

determining, for each choice word, a passage count representative of the number of context words having a passage that includes a word that matches the choice word and generating each said correlation signal, representative of the likelihood that said choice word is a match for the ambiguous word, as a function of said passage count.

21. A method according to claim 20 wherein said step of generating said correlation signal includes the further step of

determining a frequency count for each choice word representative of the number of times the choice word occurs in the passages, and

generating each said correlation signal as function of said passage count weighted by said frequency count.

22. A method according to claim 13 wherein said step of generating correlation signal includes the further steps of

generating an occurrence signal for each of said choice words as a function of an n-gram language model and one or more of said context words to represent the likelihood that said choice word would be selected from a vocabulary of words, and

combining said occurrence signal with said correlation signal, representative of the likelihood that said choice word is a match for the ambiguous word, to adjust said correlation signal as a function of said likelihood of occurrence for said choice word.

23. A method according to claim 13 wherein the step of generating said correlation signal includes the step of

adjusting said correlation signal as a function of a word count signal representative of the number of words separating the ambiguous word from a context word having a passage that contains a word that matches the choice word.

24. A method for ascertaining an ambiguous word that occurs in a data string which includes one or more context words, comprising the steps of

generating a list of one or more choice words to represent choices for said ambiguous word,

providing a reference source having one or more passages, each passage including a relevance indicator and a series of passage words associated with the relevance indicator,

comparing a respective one of said choice words with each said relevance indicator to select one or more of said passages from said reference source.

comparing one of said context words with each said relevance indicator to select one or more of said passages from said reference source, and

generating a correlation signal for each of said respective choice words by correlating said passages of said context words with said passages of said respective choice words, to represent the likelihood that said choice word is a match for the ambiguous word.

25. A method according to claim 24 for use with a continuous speech recognition system, wherein

said step of generating said list of choice words, includes the steps of comparing select characteristics of a portion of speech to one or more word models each being representative of a word in a vocabulary, and

selecting as a function of said comparison a subset of said words from said vocabulary.

26. A method according to claim 24 wherein said step of providing a a reference source includes the step of

providing a reference source comprising a dictionary.

27. A method according to claim 24 wherein said step of providing a reference source includes the step of

providing a reference source comprising an encyclopedia.

28. A method according to claim 24 wherein the step of selecting passages from a reference source includes the steps of

comparing the choice word with each said relevance indicator, and selecting passages having a relevance indicator that contains a word that substantially matches the choice word, and

comparing a representative one of said context words with each said relevance indicator, and selecting passages associated with a relevance indicator that substantially matches the context word.

29. A method according to claim 24 wherein said step of selecting passages includes the steps of

selecting a choice word and a context word, each being in a first language,

translating said selected choice word and said context word into a second language,

accessing a reference source comprising information recorded in the second language and selecting said one or more passages therefrom, and

generating said correlation signal as a function of said selected passages recorded in the second language.

30. A method according to claim 24 wherein the step of generating said correlation signal includes the step of

determining for each choice word, a word count representative of the number of distinct words in the choice word passage that occur within a passage associated with one of the context words and generating said correlation signal as a function of said word count.

31. A method according to claim 30 wherein said step of generating said correlation signal includes the further step of

determining a frequency count for each passage word representative of the number of times the passage word occurs in a passage, and

generating said correlation signal as function of said word count weighted by said frequency count.

32. A method according to claim 24 wherein said step of generating a correlation signal includes the further steps of

generating an occurrence signal for each of said choice words as a function of an n-gram language model and one or more of said context words to represent the likelihood that said choice word would be selected from a set of vocabulary words, and

combining said occurrence signal with said correlation signal to adjust said correlation signal as a function of said likelihood of occurrence for said choice word.

33. A method according to claim 24 wherein said step of generating said correlation signal includes the step of

adjusting said correlation signal as a function of a word count signal representative of the number of words separating one of the context words from the ambiguous word being recognized.

34. A method according to claim 24 wherein the step of generating the correlation signal includes the step of comparing the selected passages associated with a context word and the selected passages associated with a choice word to detect a matching n-gram.

35. A method for ascertaining an ambiguous phrase that occurs in a data string which includes at least one context word, comprising the steps of

generating a list of one or more choice phrases, each of which represents one or more words, to represent choices for the ambiguous phrase,

providing a reference source having one or more passages, each passage including a relevance indicator and a series of words associated with the relevance indicator,

selecting as a function of one or said words in said choice phrase and each said relevance indicator, at least one of said passages from said reference source,

generating a correlation signal, as a function of said selected passage, and at least one of said context words, to represent the likelihood that said choice phrase is a match for the ambiguous phrase, and

selecting, as a function of said correlation signal, at least one of said choice phrases to represent the ambiguous phrase being identified.

36. A method according to claim 35 wherein the step of selecting passages includes the step of

selecting a passage for each of a plurality of the words in the choice phrase.

37. A method according to claim 35 wherein the step of generating a correlation signal includes the step of

generating for each choice phrase a word count signal representative of the number of words in the selected passages that match the words in the respective choice phrase.

38. A method according to claim 35 wherein the step of selecting passages from a reference source includes the step of

comparing at least one word of the choice phrase with the relevance indicators, and selecting passages having a relevance indicator that contains a word that substantially matches said word of the choice phrase.

39. Apparatus for recognizing an ambiguous word that occurs in a data string which includes at least one context word, comprising

means for providing a list of one or more choice words to represent choices for said ambiguous word,

means for selecting a choice word and for accessing a reference source having one or more passages, each passage including a relevance indicator and a series of words associated with the relevance indicator, to provide a list of passage words having a known association with said selected choice word,

means for generating a correlation signal for each of said choice words as a function of said passage words, to represent the likelihood that said choice word is substantially similar to said ambiguous word, and

means for selecting, as a function of said correlation signal, at least one of said choice words to represent said ambiguous word.

40. Apparatus according to claim 39 wherein said means for providing a list of passage words comprises

means for selecting one or more of said passage words from an electronically addressable reference source having information related to said selected choice word.

41. Apparatus according to claim 39 wherein said choice word is in a first language and said reference source contains passages in a second language, further including

translating means for translating said selected choice word into the second language to provide a list of passage words in the second language.

42. In an apparatus that employs context for recognizing an ambiguous word that occurs in a data string which includes one or more context words, the improvement comprising

means for selecting a list of one or more choice words to represent choices for said ambiguous word, means for accessing a reference source, having one or more each passage including a relevance indicator and a series of words associated with the relevance indicator, to provide a list of passage words having a known association with said selected word, and

means for selecting, as a function of each relevance indicator, at least one of said choice words to represent said ambiguous word.

43. Apparatus for recognizing an ambiguous word that occurs in a data string which includes at least one context word, comprising

means for providing a list of one or more choice words to represent choices for said ambiguous word,

means for selecting said context word and for accessing a reference source, having one or more passages, each passage including a relevance indicator and a series of words associated with the relevance indicator, to provide a list of passage words having a known association with said selected context word,

means for generating a correlation signal for each of said choice words as a function of said passage words, to represent the likelihood that said choice word is substantially similar to said ambiguous word, and

means for selecting, as a function of said correlation signal, at least one of said choice words to represent said ambiguous word.